THE REAL EFFECTS OF LIQUIDITY DURING THE FINANCIAL … · General Motors Acceptance Corporation (GMAC), the ﬁnancing arm of GM and one of the world’s largest providers of auto

THE REAL EFFECTS OF LIQUIDITY DURING THEFINANCIAL CRISIS: EVIDENCE FROM AUTOMOBILES∗

EFRAIM BENMELECH

RALF R. MEISENZAHL

RODNEY RAMCHARAN

Illiquidity in short-term credit markets during the financial crisis mighthave severely curtailed the supply of nonbank consumer credit. Using a new dataset linking every car sold in the United States to the credit supplier involvedin each transaction, we find that the collapse of the asset-backed commercialpaper market reduced the financing capacity of such nonbank lenders as captiveleasing companies in the automobile industry. As a result, car sales in countiesthat traditionally depended on nonbank lenders declined sharply. Although otherlenders increased their supply of credit, the net aggregate effect of illiquidity oncar sales is large and negative. We conclude that the decline in auto sales duringthe financial crisis was caused in part by a credit supply shock driven by theilliquidity of the most important providers of consumer finance in the auto loanmarket. These results also imply that interventions aimed at arresting illiquidityin short-term credit markets might have helped contain the real effects of thecrisis. JEL Codes: G01, G23, L62.

∗We thank Bo Becker, Gadi Barlevi, Gabriel Chodorow-Reich, Dan Covitz,Diana Hancock, Arvind Krishnamurthy, Gregor Matvos, Jonathan Parker, WaynePassmore, Karen Pence, Phillip Schnabl, Andrei Shleifer (the editor), JeremyStein, Philip Strahan, Amir Sufi, three anonymous referees, and seminar par-ticipants at the 2015 AEA Meetings, Basel Research Task Force, Berkeley (Haas),CSEF-IGIER Symposium in Economics and Institutions (Capri), Cornell Univer-sity, Dutch National Bank, Emory University, Entrepreneurship and Finance Con-ference in memory of Ola Bengtsson (Lund University), Federal Reserve Board,Federal Reserve Bank of Chicago, Federal Reserve Day Ahead Conference, George-town University, Georgia State University, Hong Kong University, Indiana Uni-versity (Kelley School of Business), NBER Summer Institute, NBER CorporateFinance Meeting, NBER Monetary Economics Meeting, Northwestern Univer-sity (Kellogg), Pennsylvania State University (Smeal), Singapore ManagementUniversity, Stanford University, Tsinghua University, Tulane University, Univer-sity of Munich, and University of Illinois at Urbana-Champaign for very helpfulcomments. Della Cummings, Sam Houskeeper, Jeremy Oldfather, and JeremyTrubnick provided excellent research assistance. Benmelech is grateful for finan-cial support from the National Science Foundation under CAREER award SES-0847392. The views expressed here are those of the authors and do not necessarilyreflect the views of the Board of Governors or the staff of the Federal ReserveSystem.

C© The Author(s) 2016. Published by Oxford University Press, on behalf of the Presi-dent and Fellows of Harvard College. All rights reserved. For Permissions, please email:[email protected] Quarterly Journal of Economics (2017), 317–365. doi:10.1093/qje/qjw031.Advance Access publication on November 13, 2016.

317

mailto:[email protected]

318 QUARTERLY JOURNAL OF ECONOMICS

I. INTRODUCTION

Financial crises can have large adverse effects on real eco-nomic activity. Illiquidity in one corner of the financial systemand large realized balance sheet losses in the financial sector canlessen the aggregate credit supply and spur economic decline.1

Consistent with these theoretical predictions, there is growing ev-idence from the 2007–2009 financial crisis that the balance sheetlosses incurred by traditional financial institutions—banks andcredit unions—might have led to a fundamental postcrisis disrup-tion in credit intermediation, contributing to the recession andthe subsequent slow economic recovery (Chodorow-Reich 2014;Ramcharan, van den Heuvel, and Verani forthcoming).2

However, nonbank financial institutions—such as financeand leasing companies—are historically important sources ofcredit, especially for purchases of such consumer durable goods asautomobiles and appliances (Ludvigson 1998). Before the crisis,for example, nonbank lenders financed more than half of all newcars bought in the United States (Online Appendix Table A.1).Unlike most traditional banks, nonbank financial institutionshave connections to the shadow banking system, relying for fund-ing primarily on short-term markets, such as the asset-backedcommercial paper (ABCP) market.

We investigate how runs in the ABCP market and the lossof financing capacity at nonbank institutions, such as the captiveleasing arms of car manufacturers, might have curtailed the sup-ply of auto credit, led to the collapse in car sales, and exacerbatedthe financial difficulties of companies such as General Motors(GM) and Chrysler that were already verging on bankruptcy. Be-tween 2007 and 2008, short-term funding markets in the UnitedStates dried up as money market funds (MMFs) and other tra-ditional buyers of short-term debt fled these markets (Covitz,Liang, and Suarez 2013). Although the initial decline in 2007 wasdriven mainly by ABCP backed by mortgage-backed securities,the decline following the Lehman Brothers bankruptcy affectedall ABCP issuers.

By early 2009, growing illiquidity in the ABCP market—a keysource of short-term credit in the United States—made it difficult

1. See, for example, Diamond and Rajan (2005, 2011), Shleifer and Vishny(2010).

2. The crisis may have also disrupted intermediation even at such nontradi-tional lenders as Internet banks (Ramcharan and Crowe 2013).

https://academic.oup.com/qje

LIQUIDITY DURING THE FINANCIAL CRISIS 319

for many nonbank intermediaries to roll over debt or secure newfunding (Campbell et al. 2011). This illiquidity coincided with thecollapse of several large nonbank lenders, chief among them theGeneral Motors Acceptance Corporation (GMAC), the financingarm of GM and one of the world’s largest providers of auto financ-ing. At the same time, automobile sales fell dramatically in 2008and 2009, and GM and Chrysler eventually filed for Chapter 11bankruptcy protection.

To uncover the economic consequences of these disruptions inshort-term funding markets, we use a proprietary micro-level dataset from Polk automotive data from HIS Global (Polk) of all newcar sales in the United States. Our data set matches every new carsale to its financing source (e.g., auto loan or lease) and identifiesthe financial institution involved in the sale. The data, reportedquarterly starting in 2002, also identify each vehicle’s make andmodel along with county of registration. This micro-level detailedinformation and its spatial nature enable us to develop an em-pirical identification strategy to identify how the loss of financingcapacity in the shadow banking system might have affected U.S.car sales.

Our identification strategy hinges on the notion that by theend of 2008, liquidity runs in the ABCP market and dislocationsin other short-term funding markets might have curtailed thefinancing capacity of nonbank institutions, notably the captive fi-nancing arms of automakers. We show cross-sectionally that incounties where buyers are historically more dependent on non-bank lenders for auto credit, sales of cars fell even more sharplyin 2009. In particular, a 1 standard deviation increase in nonbankdependence is associated with a 1 percentage point or 0.08 stan-dard deviation decline in the growth in new car transactions overthe 2008–2009 period. This point estimate implies that even withthe unprecedented interventions aimed at unfreezing short-termfunding markets in 2008 and 2009, as well as the bailout of U.S.automakers and their financing arms, the liquidity shock to non-bank financing capacity might explain about 31% of the drop incar sales in 2009 relative to 2008.

Nonbank lenders tend to serve lower-credit-qualityborrowers—the very people identified as most affected bythe Great Recession. There is compelling evidence that theseborrowers suffered significantly from the collapse in house prices,reducing their demand for automobiles and other durable goods(Mian and Sufi 2014a). These borrowers are also more likely to


have faced a reduction in their credit card limits. Rather thanreflecting the effects of diminished captive financing caused byilliquidity in short-term funding markets, the results reportedhere could reflect a more general contraction in credit to riskierborrowers.

To address this challenge to causal inference, we show thatour county-level results are robust to the inclusion of the mostcommon proxies for household demand—house prices, householdleverage, and household net worth—as well as to measures of un-employment (Mian and Sufi 2014a). We also find evidence of sub-stitution: sales financed by noncaptive lenders—those financialinstitutions more dependent on traditional deposits for funding—actually rose during this period in counties where borrowers hada higher dependency on nonbank credit. The evidence on substi-tution from nonbank to bank financing suggests that our resultsare unlikely to be driven by latent demand factors; rather, theyreflect a credit supply shock.

The rich data, especially the make-segment information, al-low us to address other concerns about county-level omitted vari-ables. Within the same make, manufacturers use different modelsto appeal to diverse consumers at varying price points. GM, forexample, markets Chevrolets to nonluxury buyers, whereas it pro-motes Cadillacs to wealthier consumers. The effects of the GreatRecession on likely buyers of Chevrolets were probably very differ-ent than on potential buyers of Cadillacs, even for those living inthe same county. We can thus use county-segment fixed effectsto nonparametrically control for differences in demand withina county across different model segments. Our results remainunchanged.

Whereas the Polk data set is rich in its coverage of informationregarding automobiles, it contains no information on borrowercharacteristics. We supplement the data from Polk with a largemicro-level panel data set from Equifax of about three millionindividuals. The Equifax data include the dynamic risk score ofeach borrower, along with age, automotive credit, mortgage, andother credit usage measures. For automotive debt, the data setalso identifies whether credit was obtained from a nonbank orbank lender. Although Equifax does not provide as rich a set ofinformation about the car purchase as does Polk, it has a wealthof borrower characteristics that directly address concerns aboutborrower credit quality, credit access, and latent demand amongusers of nonbank credit relative to other sources of automotive


credit. Combining information from Polk and Equifax enables usto alleviate concerns pertaining to omitted variables at both theborrower and the car level.

Using the Equifax data and controlling for borrower’s riskscore, homeownership status, and other observables, we find sig-nificant evidence that for borrowers living in counties more tra-ditionally dependent on nonbank financing, the probability of ob-taining nonbank credit fell sharply over the 2008–2009 period,falling to zero in late 2009. Falsification tests reveal no similarpattern for either mortgage or revolving lines of credit. Further-more, we find that access to nonbank automotive credit declinedsharply toward the end of 2008 and again in the second half of2009, even among borrowers with high credit scores.

Taken together, these results imply that funding disruptionsin short-term credit markets during the financial crisis signif-icantly diminished car sales. This evidence of a credit supplyshock adds to our understanding of financial crises more broadlyand complements those papers that emphasize alternative mech-anisms, such as the role of debt, deleveraging, and regulation,that might shape post–credit boom economies (see Mian and Sufi2010, 2014a; Mian, Rao, and Sufi 2013; Rajan and Ramcharan2015, forthcoming). We argue that a credit supply channel duringthe crisis was especially important in the new car auto marketbecause more than 80% of new cars in the United States are fi-nanced through leases and auto loans from nonbank and banklenders; under 20% are bought in all-cash transactions. Our ev-idence also tentatively suggests that the Treasury and FederalReserve programs aimed at stopping illiquidity in credit marketsmight have helped contain the real effects of the crisis (Goolsbeeand Krueger 2015).

This article adds to the broader literature on the effects offinancial markets and bank lending on real economic outcomes.3

But whereas previous studies of the financial crisis document theimportance of short-term funding for banks’ liquidity and lend-ing, less is known about the consequences of the collapse of short-term funding markets. Also less well understood is the importanceof leasing companies and nonbank institutions in the provisionof credit in auto markets and how these institutions might be

3. See Khwaja and Mian (2008); Brunnermeier (2009); Gorton (2010);Ivanshina and Scharfstein (2010); Acharya, Schnabl, and Suarez (2011); Cornettet al. (2011); Gorton and Metrick (2012); Acharya and Mora (2013); and Beckerand Ivashina (2014).


connected to nontraditional sources of financing. We fill this voidby documenting that the collapse of short-term funding reducedauto lending by financial institutions, which in turn resulted infewer purchases of cars and reduced economic activity. We alsoprovide evidence that, because the ABCP market collapse cur-tailed the financing capacity of many captive financing companies,illiquidity in the short-term funding markets might have playedan important role in limiting the supply of nonbank consumercredit during the crisis.4

The rest of the article is organized as follows. Section II de-scribes the institutional background of captive leasing and institu-tions’ reliance on ABCP funding. Section III presents the data andthe main summary statistics. Section IV describes the construc-tion of our measure of captive dependence. Section V displays theempirical results on the collapse of auto sales using the Polk data.Section VI presents our micro-level analysis using the Equifaxdata. Section VII concludes.

II. AUTOMOTIVE NONBANK CREDIT

Nonbank financial institutions, especially the captive financ-ing arms of the major automobile manufacturers, have long beenimportant suppliers of automotive credit in the United States.Online Appendix Table A.1 shows that in 2005 about half of auto-motive credit came from nonbank sources of credit. Among thesenonbank purveyors of credit, captives accounted for around 90%of nonbank financing in 2008 (Online Appendix Table A.2). Therise of nonbank automotive financing, especially that of captives,arose because the automobile industry’s unique combination ofhigh cost, mass appeal, and independent dealership networks re-quired a new form of financing to expand distribution and sales.A further impetus came from the reluctance of many commercialbanks to use cars as collateral.

Banks were reluctant to make car loans because cars were arelatively novel and difficult-to-value durable good. For example,banks had scant information about a model’s depreciation path,especially given that the introduction of new models often led to asharp drop in the resale value of outgoing models. When banks did

4. Kacperczyk and Schnabl (2010) provide a detailed account of the collapseof the commercial paper market during the financial crisis of 2007–2009. Gao andYun (2013) study the consequences of illiquidity in the commercial paper marketfor corporate borrowing.




make car loans, interest rates often approached the legal maxi-mum. Some bankers even thought it unwise for commercial banksto provide credit for a luxury good out of concern that such creditmight discourage the virtue of thrift (Phelps 1952). Last, car saleswere highly seasonal, and banks’ reluctance to provide automotivefinancing affected the ability of dealers to finance their inventories(Hyman 2011).

The organizational form of captives was a response to thesefrictions. Captives such as GMAC, founded in 1919, were verticallyintegrated into the manufacturer and better able to overcome in-formational frictions surrounding the value of car collateral.5 Forexample, they knew the model release schedule well ahead ofarm’s-length lenders. Vertically integrated captives were also lessencumbered by moral objections to consumer spending on cars. Onthe dealer side of the transaction, captives often allowed the dealerto intermediate captive credit and earn additional markups. Cap-tives became important sources of credit or floorplan financing forthe dealer—a form of credit collateralized by the dealer’s auto in-ventory.6 Captives thus relaxed financial constraints at both thedealership and consumer sides of the sales transaction.7

The modern auto credit market is large because most newcars in the United States are bought on credit through eithercar loans or leasing. At its peak in 2006, auto credit was $785billion, accounting for 32% of consumer debt, and assets atGMAC, then the largest of the captive financiers, totaled around$26 billion. Captive lessors are often seen as providing creditto riskier borrowers (Barron, Chong, and Staten 2008; Einav,

5. Import brands such as Toyota tend to rely more heavily on nonbank captivesthat are not vertically integrated. For example, existing nonbank lenders, such asWorld Omni Financial, created a dedicated subsidiary, Southeast Toyota Finance,in 1981, to help Toyota establish a foothold in the North American market in keygeographic regions. Toyota Motor Credit was established only in 1982 and focuseson markets outside the Southeast (Kaisha 1988).

6. These points are echoed by William C. Durant in announcing the formationof GMAC in a letter dated March 15, 1919: “The magnitude of the business haspresented new problems in financing which the present banking facilities seem notto be elastic enough to overcome. . . . This fact leads us to the conclusion that theGeneral Motors Corporation should lend its help to solve these problems. Hencethe creation of General Motors Acceptance Corporation; and the function of thatCompany will be to supplement the local sources of accommodation to such extentas may be necessary to permit the fullest development of our dealers’ business”(Sloan 1964, p. 303).

7. Murfin and Pratt (2015) expand on these ideas within a theoretical modeland provide evidence based on equipment leasing.


Jenkins, and Levin 2013).8 In 2006, the median FICO score forcar buyers obtaining nonbank credit was 640, as opposed to 715for buyers using bank credit.

Before the financial crisis, securitization gave nonbank sup-pliers of automotive credit new ways to tap into cheap fund-ing (Calder 1999; Hyman 2011).9 In particular, ABCP becamenonbank lenders’ main source of funding, enabling them to turnrelatively illiquid auto term loans into liquid assets that could beused to obtain funding for new loans. In this form of securitization,pooled auto loans are placed in a special-purpose vehicle (SPV)that is bankruptcy remote from the originating captive lessor. TheSPV then issues short-term secured commercial paper (ABCP) tofinance loans and market the commercial paper—generally witha duration of no more than three months (see Acharya, Schnabl,and Suarez 2011 for a detailed discussion of ABCP structures).

MMFs and other institutional investors seeking to investin liquid and high-yield short-term assets are the main buyersof commercial paper. In mid-2007, just before the turbulence incredit markets, MMFs held about 40% of outstanding commercialpaper in the United States. The bankruptcy of Lehman Broth-ers on September 15, 2008, and the “breaking of the buck” atReserve Primary Fund the next day triggered heavy outflowsfrom MMFs, leading the Treasury to announce an unprecedentedguarantee program for virtually all MMF shares. The Federal Re-serve followed suit by announcing a program to finance purchasesof ABCP—which were highly illiquid at the time—from MMFs.Despite these interventions, flows into MMFs remained highly

8. Charles, Hurst, and Stephens (2010) document that minorities, particularlyAfrican Americans, are more likely to receive auto loans from financing companiesand to pay, on average, higher interest rates on those loans. One plausible ex-planation for this pattern is that minorities have, on average, lower credit scoresand therefore are more likely to receive financing from captives. For a detailedanalysis of subprime auto-lending contracts, see Adams, Einav, and Levin (2009);and Einav, Jenkins, and Levin (2012).

9. Online Appendix Table A.3, based on nonpublic data collected by the FederalReserve, demonstrates the importance of commercial paper as a source of fundingfor selected major automobile captives active in the United States. Given thenature of the data, we cannot disclose the identities of the captive lessors in thetable and instead label them Captive 1 through Captive 4. As Online AppendixTable A.3 shows, commercial paper was a major source of funding for three out ofthe four captive lessors. Although commercial paper accounted for just 10.2% ofone lessor’s liabilities (Captive 3), the other three relied much more heavily on thisform of short-term funding, with the share of commercial paper in their liabilitiesranging from 45.9% (Captive 2) to 75.1% (Captive 4).




erratic, and MMFs significantly retrenched their commercial pa-per holdings. In the three weeks following Lehman’s bankruptcy,prime MMFs reduced their holdings of commercial paper by$202 billion, a steep decline of 29%.

The reduction in commercial paper held by MMFs led to asharp rise in borrowing costs for issuers of commercial paper.ABCP issuances also fell sharply amid the turmoil in short-termcredit markets, and the sharp outflows of assets from MMFs in thethird quarter of 2008 precipitated a run on many of these auto-related securitization pools. Online Appendix Figure A.1 displaysthe outstanding amount of ABCP issued by SPVs associated withthe captive leasing arms of the big three U.S. automakers: GMAC,Chrysler Financial (CF), and Ford Motor Credit (FMC). Althoughthe ABCP market began to weaken in 2007, automakers’ issuanceof ABCP began to collapse in the third quarter of 2008. Together,the big three captive lessors had about $40 billion worth of ABCPoutstanding in 2006 before they largely collapsed by the end of2009.10

Before turning to the data and statistical tests, we providenarrative-based evidence on how the decline in nonbank financingcapacity might have affected automobile sales. Although nonbanklenders such as captives are key providers of consumer credit, theyare also an important source of credit for auto dealerships. Thefloorplan financing provided by nonbank lenders enables dealer-ships to purchase their car inventory. Although it is not easy toobtain dealership-level data on floorplan loans, we have read thefinancial reports of the largest publicly traded automotive deal-erships in the United States to understand the challenges thatdealerships faced during the Great Recession.

In these reports, these dealerships frequently list a lack offinancing for both consumers and dealerships as a first-order rea-son for the decline in auto sales. Collectively, these reports pointto the possibility that the illiquidity of nonbank lenders might

10. Ford’s financing arm, FMC, survived the crisis in part because of its con-tinued access to the Federal Reserve’s Commercial Paper Funding Facility (CPFF),which bought ABCP to alleviate liquidity pressures in the funding markets afterthe Lehman Brothers collapse. The Federal Reserve announced the CPFF to pro-vide a liquidity backstop for U.S. commercial paper issuers with high short-termcredit ratings on October 14, 2008. Before losing access in January 2009, GMACheavily relied on CPFF, selling a total of $13.5 billion in ABCP to the facility. Incontrast to GMAC and CF, FMC was able to maintain its short-term credit ratingand never lost access to CPFF, from which it had raised almost $16 billion bysummer 2009 and then began to raise funds from private investors.



have led to a decline in auto sales through a credit supply channelthat affected not only consumers but also car dealerships. We col-lect and reproduce discussions from the Form 10-Ks of the largestpublicly traded dealership companies in the United States thatpertain to the role of nonbank credit and the automotive industryin general and during the financial crisis and report on them inthe Online Appendix.

III. DATA AND SUMMARY STATISTICS

For our county-level analysis, we use a proprietary data setfrom R. L. Polk & Company that records all new car sales in theUnited States. Beginning in 2002, for each new car purchasedin the United States, the data set identifies the vehicle makeand model, such as Ford (make) Focus (model) or Toyota (make)Camry (model), and whether the car was purchased by a privateconsumer (retail purchase), a firm (commercial purchase), or thegovernment. The data set also details the county, year, and quarterof vehicle registration. Because we are interested in identifyingthe effect of a credit supply shock on household consumption, wefocus exclusively on retail purchases. Moreover, for each retailcredit transaction starting in the first quarter of 2008, Polk liststhe name of the financial institution and type of financial servicesprovider: bank, credit union, or nonbanks such as the automaker’scaptive financing arm.

Using the Polk data, we replicate the well-known observa-tion that durable goods purchases—particularly automobiles—declined sharply during and after the financial crisis. Online Ap-pendix Figure A.2 plots the number of automobiles sold annuallyfrom 2002 to 2013. Car sales plummeted from a peak of 17 millionunits in 2006 to 11 million units in 2009 before rebounding slightlyin 2010 and 2011. In 2012, auto sales had recovered to around14 million units sold, and by 2013 sales approached precrisis lev-els. The decline in automobiles sales during the crisis was drivenlargely by retail auto sales (see Online Appendix Figure A.3).

Summary statistics of annual county-level retail auto salesare reported in the Online Appendix (Table A.4). County-levelmean sales dropped from 3,866 units in 2007 to 3,168 and 2,563,respectively, in 2008 and 2009, reflecting the dramatic decline inauto sales during the crisis. This pattern of dramatic decline is notdriven by outlier counties and can also be observed by inspectingsuch sample order statistics as the median and the first and thirdquartiles. Figure I displays the spatial variation in the collapse of







FIG

UR

EI

Cou

nty

-Lev

elC

han

gein

Ret

ailC

arS

ales

,200

8–20

09

Ret

ailc

arsa

les

are

the

sum

ofre

tail

leas

esan

dre

tail

purc

has

esin

Pol

k.


retail car sales, defined as the percentage change in retail auto-mobile sales from 2008 to 2009 within a county. Counties in NewEngland and parts of the upper West experienced a smaller dropin retail auto sales relative to the majority of counties in the Southand West.

IV. MEASURING NONBANK DEPENDENCE

To analyze the role of nonbank financing in the collapse of carsales, we construct a measure of a county’s dependence on non-bank financing. We define our measure of nonbank dependence asthe ratio of the number of retail auto sales financed by nonbanks tothe number of all retail financed transactions in the county in thefirst quarter of 2008. An alternative definition of this measure is todivide the number of retail auto sales financed by nonbanks by allretail transactions in the county, regardless of whether they arefinanced. We focus on the former measure because it alleviatesthe concern that nonbank dependence proxies for more generalcredit usage and demand in a county. To wit, when dependence isdefined as a share of total retail sales, it might be high in countiesthat use more financing, which might be correlated with demandshocks. By using nonbank-financed transactions as a share of allretail financed transactions in the county, we purge the pure fi-nancing effect and focus instead on the intensive margin of thecomposition of financing.

The timing of our baseline measure of nonbank dependencecan also affect inference. The first quarter of 2008 is the earli-est available date that Polk records lender information. But ifdealers and consumers began substituting away from nonbank fi-nancing to other lenders during this period, the baseline measuremight already reflect the effects of this substitution, rather than acounty’s historic dependence on nonbank credit. Also, because thebaseline dependence measure is based on first quarter 2008 data,seasonality in the provision of credit across lenders could lead toinaccurate estimates of a county’s nonbank dependence.

These measurement concerns are valid. But they are likely tobe mitigated by the relationship-based nature of nonbank autocredit, most of which consists of captive credit. Captive rela-tionships are especially strong at the wholesale or dealershiplevel and can render the cross-county variation in nonbank de-pendence persistent, at least before the full onset of the finan-cial crisis. Figure II plots the county-level variation in nonbank


FIG

UR

EII

Cou

nty

-Lev

elS

har

eof

Ret

ailC

ars

Fin

ance

dby

Non

ban

ksin

2008

Q1

Ret

ail

car

sale

sar

eth

esu

mof

reta

ille

ases

and

reta

ilpu

rch

ases

inP

olk.

Th

esh

are

isde

fin

edre

lati

veto

all

reta

iltr

ansa

ctio

ns

inth

eco

un

ty,r

egar

dles

sof

the

sou

rce

offi

nan

cin

g.


dependence, measured in the first quarter of 2008. Not surpris-ingly, Michigan—the headquarters of the three major domesticmanufacturers and their respective captive financing arms—hasthe largest share of nonbank-financed transactions in the UnitedStates. In areas where other auto manufacturers have a long-standing presence and dealers have close relationships with cap-tives, such as Alabama and Tennessee, captives also appear todominate credit transactions.

To address these measurement concerns, we use data fromEquifax to supplement our Polk-based baseline county-level non-bank dependence measure. Equifax, one of the three major creditbureaus in the United States, collects data on individuals’ li-abilities, including their car purchases; in the version of thedata set available to us, it identifies whether the source ofautomotive credit is a nonbank financier along with the borrower’szip code. These data are available quarterly and extend back to2006, which enables us to construct measures of nonbank depen-dence at least two years before the onset of the financial crisis.11

We draw a 10% random sample from Equifax, which yields a panelof about three million households. As Figure III demonstrates, thequarterly growth in car sales derived from both Polk and Equifaxare very similar.

We aggregate the Equifax data at the county level and createtwo measures of nonbank dependence. These measures are: (i) theratio of nonbank financed transactions to all financed transactionsin the county in the first quarter of 2008, which corresponds tothe time period in the baseline Polk measure, and (ii) the ratio ofnonbank financed transactions to all finance transactions in 2006.Table I reports the summary statistics for the two Equifax-basedmeasures of nonbank dependence (columns (1) and (2)); the Polkderived ratio of nonbank to all new transactions (column (3)); andthe baseline ratio of nonbank to all financed transactions, alsofrom Polk (column (4)), along with key control variables.

The basic summary statistics suggest that nonbanks accountfor about 40% of all auto purchases (column (3)) and for about 52%of all financed purchases (column (4)). The dependence measuresderived from Equifax and Polk are very similar to each other, al-though the average incidence of nonbank credit appears to be alittle smaller in 2006 compared with that observed in the firstquarter of 2008. The cross-sectional variation in all four variables

11. Equifax does not list the name of the credit supplier.


FIGURE III

Quarterly Growth in New Car Sales—Comparing Polk and Equifax

The figure plots the quarter on quarter growth in car sales, as reported byPolk and Equifax.

is very similar. Online Appendix Table A.5 reports the coefficientfrom regressing separately the Equifax 2008 first quarter mea-sure of dependence separately on the other three alternative de-pendence variables, controlling for state fixed effects. The pointestimates in these regressions range from 0.762 to 0.821 and areall statistically significant at the 1% level. In robustness tests wepresent later, we show that our baseline estimates are relativelyunchanged across the alternative measures of captive dependence.

IV.A. The Determinants of Nonbank Dependence

To understand the determinants of a county’s dependenceon nonbank financing, we estimate cross-county regressions ofnonbank dependence on a number of county-level demographicvariables and report the results in Table II. The first 12 columns ofTable II show the coefficient from univariate regressions of non-bank dependence on each of these variables. Counties more depen-dent on nonbank credit are generally more populous, have highermedian income, and have greater income inequality (measured



TA

BL

EI

SU

MM

AR

YS

TA

TIS

TIC

S

Cap

tive

depe

nde

nce

Equ

ifax

2008

Q1

Equ

ifax

2006

Pol

k20

08Q

1

Pol

k20

08Q

1fi

nan

ced

Cou

nty

area

,log

Pop

ula

tion

,lo

gM

edia

nin

com

e,lo

g

Afr

ican

Am

eric

anpo

pula

tion

,lo

g

Wh

ite

popu

lati

on,

log

Gin

ico

effi

cien

t

Em

ploy

men

tin

auto

mob

ile

sect

or,s

har

e

Med

ian

cred

itsc

ore,

2008

Q1

(Tra

ns

Un

ion

)

Mea

n0.

490.

450.

390.

526.

4710

.25

0.66

6.78

10.0

60.

430.

4267

5.74

Med

ian

0.5

0.45

0.38

0.52

6.42

10.1

510

.65

6.95

9.99

0.43

067

925

thpe

rcen

tile

0.38

0.35

0.32

0.44

6.06

9.32

10.5

4.9

9.12

0.41

063

475

thpe

rcen

tile

0.6

0.56

0.45

0.6

6.82

11.0

910

.79

8.73

10.9

40.

450.

0471

7M

in0.

10

0.08

0.14

0.69

4.39

9.86

04.

170.

210

507.

5M

ax0.

881

11

9.91

16.1

11.6

514

.11

15.4

20.

6418

.66

811

Sta

nda

rdde

viat

ion

0.15

0.18

0.1

0.12

0.87

1.45

0.25

2.61

1.44

0.04

1.44

54.6

6

Not

es.T

his

tabl

epr

esen

tsth

esu

mm

ary

stat

isti

csfo

rco

un

tych

arac

teri

stic

su

sed

inth

eem

piri

cal

anal

ysis

.In

colu

mn

(1),

non

ban

kde

pen

den

ceis

the

rati

oof

non

ban

k-fi

nan

ced

tran

sact

ion

sto

all

fin

ance

dtr

ansa

ctio

ns

ina

cou

nty

asof

2008

Q1

and

repo

rted

inE

quif

ax.C

olu

mn

(2)

defi

nes

non

ban

kde

pen

den

cesi

mil

arly

,bu

tta

ken

over

all

of20

06.C

olu

mn

(3)

defi

nes

non

ban

kde

pen

den

ceas

the

rati

oof

non

ban

k-fi

nan

ced

tran

sact

ion

sto

all

sale

sin

the

cou

nty

,in

clu

din

gse

lf-fi

nan

ced

tran

sact

ion

s,as

of20

08Q

1an

dre

port

edin

Pol

k.C

olu

mn

(4)d

efin

esn

onba

nk

depe

nde

nce

asth

era

tio

ofn

onba

nk-

fin

ance

dtr

ansa

ctio

ns

toal

lfin

ance

dtr

ansa

ctio

ns

ina

cou

nty

asof

2008

Q1

and

repo

rted

inP

olk.

Pop

ula

tion

,cou

nty

area

,med

ian

hou

seh

old

inco

me,

Gin

ico

effi

cien

t,po

vert

yra

te,A

fric

anA

mer

ican

popu

lati

on,a

nd

wh

ite

popu

lati

onar

eta

ken

from

the

Am

eric

anC

omm

un

ity

Su

rvey

.Em

ploy

ees

inau

tom

obil

ese

ctor

and

tota

lem

ploy

men

tar

eta

ken

from

the

Qu

arte

rly

Cen

sus

ofE

mpl

oym

ent

and

Wag

es(Q

CE

W).


TA

BL

EII

TH

ED

ET

ER

MIN

AN

TS

OF

NO

NB

AN

KD

EP

EN

DE

NC

E

Dep

ende

nt

vari

able

:non

ban

kfi

nan

ced

toal

lfin

ance

dtr

ansa

ctio

ns

(Pol

k)(1

)(2

)(3

)(4

)(5

)(6

)(7

)(8

)(9

)(1

0)(1

1)(1

2)(1

3)(1

4)(1

5)

Cou

nty

area

,−0

.004

19−0

.003

910.

0035

4lo

g(0

.003

64)

(0.0

0614

)(0

.004

71)

Pop

ula

tion

,0.

0104

∗∗∗

0.04

43∗

0.01

46lo

g(0

.002

89)

(0.0

244)

(0.0

231)

Med

ian

0.03

96∗∗

0.04

25∗∗

0.06

84∗∗

∗in

com

e,lo

g(0

.015

3)(0

.017

3)(0

.015

8)A

fric

an0.

0079

5∗∗∗

0.00

185

0.00

0138

Am

eric

anpo

pula

tion

,lo

g

(0.0

0173

)(0

.002

48)

(0.0

0227

)

Wh

ite

0.00

843∗

∗∗−0

.041

8∗−0

.009

36po

pula

tion

,lo

g(0

.002

88)

(0.0

226)

(0.0

216)

Gin

i0.

325∗

∗∗0.

300∗

∗∗0.

101

coef

fici

ent

(0.0

853)

(0.0

788)

(0.0

625)

Em

ploy

men

t−0

.046

1−0

.105

−0.2

11∗

in auto

mob

ile

sect

or,

shar

e

(0.1

60)

(0.1

45)

(0.1

14)


TA

BL

EII

( CO

NT

INU

ED

)

Dep

ende

nt

vari

able

:non

ban

kfi

nan

ced

toal

lfin

ance

dtr

ansa

ctio

ns

(Pol

k)(1

)(2

)(3

)(4

)(5

)(6

)(7

)(8

)(9

)(1

0)(1

1)(1

2)(1

3)(1

4)(1

5)

Med

ian

cred

it−0

.000

137

−4.0

6e−0

5−0

.000

692∗

∗∗sc

ore,

2008

Q1

(Tra

ns

Un

ion

)(9

.60e

−05)

(0.0

0010

3)(9

.18e

−05)

Un

empl

oym

ent

0.00

0758

−0.0

0024

40.

0012

8ra

te,2

008

(0.0

0136

)(0

.001

10)

(0.0

0089

1)U

nem

ploy

men

t−0

.001

50ra

te,2

006

(0.0

0336

)U

nem

ploy

men

t−0

.000

543

rate

,200

7(0

.002

99)

Pov

erty

rate

,0.

0006

1620

08(0

.000

710)

Ch

ange

inh

ouse

−0.0

913

pric

es,

2005

–200

6(0

.089

7)

Obs

erva

tion

s3,

092

3,09

23,

092

2,86

03,

092

3,09

23,

087

3,09

22,

534

2,53

42,

534

3,09

21,

239

2,30

12,

301

R2

0.30

90.

320

0.31

30.

371

0.31

60.

316

0.30

80.

310

0.29

90.

299

0.29

90.

309

0.42

50.

372

0.39

9

Not

es.T

his

tabl

ere

port

sth

ere

gres

sion

resu

lts

ofre

gres

sin

gco

un

ty-l

evel

non

ban

kde

pen

den

ce(P

olk)

,200

8Q

1,on

dem

ogra

phic

and

econ

omic

vari

able

sob

serv

edar

oun

dth

esa

me

peri

od.E

mpl

oym

ent

inth

eau

tom

obil

ese

ctor

isn

um

ber

ofem

ploy

ees

inth

eau

tom

obil

ese

ctor

divi

ded

byto

tale

mpl

oym

ent.

Th

eso

cioe

con

omic

vari

able

sar

eta

ken

from

the

Am

eric

anC

omm

un

ity

Su

rvey

.Cou

nty

-lev

elu

nem

ploy

men

tra

tes

com

efr

omth

eB

ure

auof

Lab

orS

tati

stic

s.E

mpl

oyee

sin

auto

mob

ile

sect

orar

eta

ken

from

the

QC

EW

.Th

ede

pen

den

tva

riab

lein

colu

mn

(15)

isn

onba

nk-

fin

ance

dto

allt

ran

sact

ion

s(P

olk)

.∗∗∗

,∗∗ ,

∗de

not

essi

gnifi

can

ceat

the

1%,5

%,a

nd

10%

leve

ls,r

espe

ctiv

ely.


by the county’s Gini coefficient). There is no evidence that eco-nomic conditions before the crisis, as proxied for by the unemploy-ment rate in either 2006 or 2007, are correlated with dependence.Nonbank dependence is also not significantly related to thehousing cycle, as measured by the change in house prices dur-ing the boom (2005–2006).

Column (14) reports results from a multivariate regressionthat includes these variables jointly. There is some evidence thatnonbank dependence is smaller in counties with more white res-idents, and the positive coefficients on both the median incomeand population variables remain significant.

Column (15) uses the extensive margin measure of nonbankdependence as the dependent variable: the number of nonbank-financed transactions divided by the number of transactions in acounty. The results confirm the concern that this extensive marginmeasure of dependence is potentially more affected by differentialcredit usage across income groups within a county; the coefficienton the median FICO score in the county is now negative andstatistically significant. Borrowers with lower FICO scores weredisproportionately affected by the crisis, however, and inferencebased on this measure of dependence might be more prone toconcerns about omitted demand-side factors.12

Although the concern about the correlation between borrowercredit quality and nonbank dependence is valid, it is important toput this concern in context. By the first quarter of 2007 only 15%of GMAC’s U.S.-serviced consumer asset portfolio was considerednonprime; GMAC was the biggest nonbank automotive lender atthat point.13 That is, the largest nonbank automotive lenders didnot concentrate on subprime borrowers, but the vast majority ofcar buyers who relied on nonbank credit were safer borrowers whohad lower sensitivity to the housing cycle.

V. NONBANK CREDIT AND THE COLLAPSE OF AUTO SALES:THE AGGREGATE EVIDENCE

Here we present county-level evidence of the relation betweennonbank credit and car sales during the crisis.

12. We are grateful to an anonymous referee for suggesting this test and theintensive measure of dependence.

13. See GMAC LLC, 8-K, April 26, 2007, File No. 001-03754. The document isavailable at the SEC’s EDGAR company filings website.


V.A. Sales Financed by Nonbank Creditors

We analyze the effect of nonbank credit dependence onthe change in automotive sales financed by nonbank creditorsbetween 2008 and 2009. We use the following baseline regressionspecification:

log(cars)2009i − log (cars)2008i = α0 + α1 × dependencei(1)

+ Xiβ + Si + ei,

where the dependent variable is the difference in the log numberof cars financed by nonbank creditors in county i between 2008and 2009. Our main explanatory variable is the county’s depen-dence on nonbank financing. The baseline county-level specifica-tions use Polk data and define dependence as the ratio of retailsales financed by nonbanks to all financed transactions in thecounty, observed in 2008 Q1—the earliest date for which the Polkdata identifies the credit provider in the transactions.14

Table III presents the results from estimating different vari-ants of the model with standard errors (in parentheses) clusteredat the state level. We also weight these county-level regressions bythe population in the county circa 2009.15 All specifications includestate fixed effects (the vector S) to absorb time-invariant state het-erogeneity. Most of our specifications also control for county-leveleconomic and demographic variables that are included in the vec-tor Xi.16 Our main coefficient of interest is α1, which measures theeffect of nonbank credit dependence on car sales during the crisis.

14. Theoretically, nonbank lenders may decide to cut funding to consumerswith no regard to the level of nonbank dependence in the county. However, inpractice they would like to maintain presence in low-dependence counties as well,which results in larger disproportionate reduction in credit in high-dependencecounties. Moreover, in the context of captive lessors, dealers (more so than con-sumers) were almost completely dependent on captives for floorplan or wholesalecredit. For example, GMAC accounted for about 85% of this wholesale credit andfor 40% of direct consumer credit. Given the nearly absolute dependence of thedealer network on GMAC credit and to protect this dealer network and ensure thelong-term viability of GM, GMAC actually shifted credit away from consumers to-ward its dealer network during the crisis, resulting in a disproportionate reductionin credit in high-dependence counties.

15. Although our results hold if we use regular OLS regressions, we weightour regressions by population to account for county size (see, e.g., Autor and Dorn2013; Mian and Sufi 2014a).

16. Table I reports summary statistics for the explanatory variables used inthese regressions.


TA

BL

EII

IN

ON

BA

NK

DE

PE

ND

EN

CE

AN

DN

ON

BA

NK

SA

LE

S

(1)

(2)

(3)

(4)

(5)

Var

iabl

es

Eco

nom

ican

dde

mog

raph

icco

ntr

ols,

wit

hou

tst

ate

fixe

def

fect

s

Eco

nom

ican

dde

mog

raph

icco

ntr

ols,

wit

hst

ate

fixe

def

fect

sU

nem

ploy

men

tan

dle

vera

geH

ouse

pric

esH

ouse

hol

dn

etw

orth

Cap

tive

depe

nde

nce

,200

8Q

1(P

olk)

,fin

ance

dtr

ansa

ctio

ns

−0.3

69∗∗

−0.3

19∗∗

∗−0

.342

∗∗∗

−0.3

48∗∗

∗−0

.316

∗∗∗

(0.1

41)

(0.0

620)

(0.0

679)

(0.0

811)

(0.0

771)

Cou

nty

area

,log

−0.0

656∗∗

−0.0

227∗∗

−0.0

243∗∗

−0.0

270∗

−0.0

271∗

(0.0

289)

(0.0

112)

(0.0

119)

(0.0

140)

(0.0

141)

Pop

ula

tion

,log

0.10

90.

108∗∗

∗0.

107∗∗

∗0.

110∗∗

∗0.

102∗∗

∗

(0.0

706)

(0.0

336)

(0.0

345)

(0.0

373)

(0.0

361)

Med

ian

inco

me,

log

−0.0

445

0.02

840.

0255

0.03

890.

0533

(0.0

773)

(0.0

344)

(0.0

347)

(0.0

434)

(0.0

428)

Afr

ican

Am

eric

anpo

pula

tion

,log

−0.0

0390

0.00

769∗

0.00

792∗

0.00

586

0.00

610

(0.0

104)

(0.0

0403

)(0

.004

39)

(0.0

0518

)(0

.004

92)

Wh

ite

popu

lati

on,l

og−0

.066

2−0

.110

∗∗∗

−0.1

10∗∗

∗−0

.106

∗∗∗

−0.0

973∗∗

∗

(0.0

606)

(0.0

286)

(0.0

301)

(0.0

315)

(0.0

310)

Gin

icoe

ffici

ent

0.51

90.

365∗∗

0.40

7∗∗0.

354∗

0.27

2(0

.348

)(0

.174

)(0

.165

)(0

.183

)(0

.201

)E

mpl

oym

ent

inau

tom

obil

e,sh

are

−1.1

36−0

.274

−0.3

03−0

.323

−0.3

96(0

.837

)(0

.261

)(0

.254

)(0

.335

)(0

.353

)M

edia

ncr

edit

scor

e,20

08Q

1(T

ran

sU

nio

n)

0.00

103∗∗

0.00

0881

∗∗∗

0.00

0851

∗∗∗

0.00

0809

∗∗∗

0.00

0842

∗∗∗

(0.0

0045

0)(0

.000

174)

(0.0

0018

4)(0

.000

236)

(0.0

0023

8)H

ouse

pric

ech

ange

0.08

12(0

.125

)


TA

BL

EII

I( C

ON

TIN

UE

D)

(1)

(2)

(3)

(4)

(5)

Var

iabl

es

Eco

nom

ican

dde

mog

raph

icco

ntr

ols,

wit

hou

tst

ate

fixe

def

fect

s

Eco

nom

ican

dde

mog

raph

icco

ntr

ols,

wit

hst

ate

fixe

def

fect

sU

nem

ploy

men

tan

dle

vera

geH

ouse

pric

esH

ouse

hol

dn

etw

orth

Un

empl

oym

ent

rate

0.00

492

0.00

363

(0.0

0438

)(0

.004

20)

Hou

seh

old

leve

rage

,200

60.

0227

0.03

63(0

.023

1)(0

.028

6)C

han

gein

hou

seh

old

net

wor

th,2

006–

2009

−0.0

437

(0.0

912)

Obs

erva

tion

s2,

825

2,82

52,

056

958

932

R2

0.23

60.

761

0.78

50.

840

0.84

2

Not

es.T

his

tabl

ere

port

sth

ere

gres

sion

resu

lts

ofes

tim

atin

geq

uat

ion

(1).

Th

ede

pen

den

tva

riab

leis

the

log

chan

gein

the

nu

mbe

rof

cars

fin

ance

dby

non

ban

ksin

2009

rela

tive

to20

08as

repo

rted

inP

olk.

Non

ban

kde

pen

den

ceis

the

rati

oof

non

ban

k-fi

nan

ced

toal

lfin

ance

dtr

ansa

ctio

ns,

(Pol

k)20

08:Q

1.P

erce

nta

geA

fric

anA

mer

ican

isth

eA

fric

anA

mer

ican

popu

lati

ondi

vide

dby

popu

lati

on.E

mpl

oym

ent

inau

tom

obil

ese

ctor

isn

um

ber

ofem

ploy

ees

inth

eau

tom

obil

ese

ctor

divi

ded

byto

tal

empl

oym

ent.

Pop

ula

tion

,cou

nty

area

,med

ian

hou

seh

old

inco

me,

Gin

ico

effi

cien

t,po

vert

yra

te,A

fric

anA

mer

ican

popu

lati

on,a

nd

wh

ite

popu

lati

onar

eta

ken

from

the

Am

eric

anC

omm

un

ity

Su

rvey

.Cou

nty

-lev

elu

nem

ploy

men

tra

tes

are

take

nfr

omth

eB

LS.

Em

ploy

ees

inau

tom

obil

ese

ctor

and

tota

lem

ploy

men

tar

eta

ken

from

the

QC

EW

.Hou

seh

old

leve

rage

isth

ede

bt-t

o-in

com

era

tio

(Fed

eral

Res

erve

Ban

kof

New

York

).H

ouse

pric

ech

ange

isth

ech

ange

inth

eh

ouse

pric

ein

dex

(Cor

eLog

ic).

Hou

seh

old

net

wor

this

from

Mia

nan

dS

ufi

(201

4a).

All

vari

able

sar

ede

fin

edin

App

endi

xA

.All

regr

essi

ons

are

wei

ghte

dby

the

cou

nty

popu

lati

onan

din

clu

dest

ate

fixe

def

fect

s,ex

cept

for

colu

mn

(1).

Sta

nda

rder

rors

are

clu

ster

edat

the

stat

ele

vel.

∗∗∗ ,

∗∗,∗

den

otes

sign

ifica

nce

atth

e1%

,5%

,an

d10

%le

vels

,res

pect

ivel

y.



Table III, column (1) presents the results of regression (1)using demographic and economic county-level controls as proxiesfor local demand, but excludes state fixed-effects. As column (1)illustrates, the nonbank dependence coefficient is economicallyand statistically significant. A 1 standard deviation increase innonbank dependence is associated with a 3.5 percentage point or0.17 standard deviation decline in the growth in nonbank financedtransactions over this period. Alternatively, moving from a countyat the 25th to the 75th percentile in nonbank credit dependenceis associated with a 5 percentage point drop in the growth innonbank financed transactions. The nonbank dependence coeffi-cient is only slightly smaller when adding state fixed effects (col-umn (2)), but in what follows all specifications use state-level fixedeffects to absorb potentially relevant regulatory, geographic, andother time-invariant state-level factors.

We control for log median income because the demand for carsmight be higher in counties with higher household income. Sim-ilarly, we control for the number of African American and whiteresidents, given the evidence that race might affect access to au-tomotive credit (Barron, Chong, and Staten 2008; Einav, Jenkins,and Levin 2013). We also add income inequality, as measuredby the Gini coefficient, as a control variable in our regressions.The majority of those who relied on nonbank credit were saferborrowers with lower sensitivity to the housing cycle. But becausenonbanks might be more likely to serve lower-credit-quality bor-rowers, who in turn might have been more exposed to the GreatRecession, we control for the median credit score in the countyusing data from Transunion. Credit scores in a county might en-dogenously respond to any credit supply disruptions, and as withthe nonbank dependence variable, our baseline specification usesthe median credit score observed in 2008 Q1. In the robustnesssection we show that these results are unchanged when usingalternative measures of borrower credit quality.

A county’s employment structure could also drive unobserveddemand shocks. In counties with strong employment links to theautomotive sector, the demand for cars might endogenously varywith the health of that sector. At the same time, these countiesmight have higher levels of nonbank dependence because of theseautomotive linkages. Figure II shows, for example, that countiesin Michigan—the headquarters of the “big three”—and countiesin states where auto manufacturers have a long-standing pres-ence (such as Alabama, Indiana, Kentucky, and Tennessee) alsohave the largest share of nonbank-financed transactions in the


United States.17 We thus add the fraction of employment in theautomotive sector as a control variable to the regression incolumns (1) and (2).

Among these demographic and economic variables, we findthat the number of African American residents in the county ispositively correlated with the number of car sales financed by non-bank lenders. Also, as one might expect, the credit quality of bor-rowers within a county is positively correlated with the growth innonbank-financed transactions. In the Online Appendix we com-bine the 2005–2009 ACS with county-level data from the 2000census to compute the changes in median income, poverty rate,overall population, and African American population within coun-ties over time. In supplementary analysis presented in OnlineAppendix Table A.6, column (1), we show that using the changesinstead of the level of these sociodemographic control variablesdoes not change the point estimate on the nonbank dependencevariable.

We next incorporate household balance sheet control vari-ables into our analysis. There is a burgeoning literature on theeffect of home prices, household leverage, and net worth on lo-cal demand and employment (see Mian and Sufi 2011, 2014a;and the broader discussion in Mian and Sufi 2014b). Someof this literature has also directly connected car purchases tohousehold-level changes in debt service (DiMaggio, Kermani, andRamcharan 2014). To the extent that our measure of nonbank de-pendence is correlated with the household balance sheet–drivendemand channel, estimates of the dependence coefficient might bebiased.

Table III, column (3) adds the 2009 county-level unemploy-ment rate as well the median debt-to-income ratio for householdsin a county in 2006 to the control variables used in columns (1)and (2).18 These data are available for a smaller subsample ofcounties, reducing the sample size from 2,825 in column (1) to2,056 counties in column (2). Yet the negative impact of nonbankdependence remains robust, with statistical significance at the 1%level and a point estimate that is very close to the one obtainedin column (1). Since unemployment and leverage might be highlycorrelated, in Online Appendix Table A.6, columns (2) and (3), we

17. Appendix A provides a detailed description of the construction of the vari-ables and their sources.

18. We thank Amir Sufi for providing debt to income ratio data.







include these variables in separate regressions; the results areunchanged.

House price dynamics was a chief catalyst behind the collapsein household demand. To address further concerns about latentdemand, column (3) directly controls for the average change inhome prices in a county from 2008 to 2009. Including this vari-able further reduces the sample size, but as Table III, column (4)demonstrates, our main finding is little changed. The house pricechange point estimate is positive, though imprecisely estimated,and suggests that a 1 standard deviation increase in house pricesis associated with a 0.05 standard deviation increase in the growthin nonbank-financed transactions. In Online Appendix Table A.6,column (4), we also include an interaction term between house-hold leverage and house price changes in the county, and our basicresults remain unchanged.

Last, in column (4) we add the change in household net worthbetween 2006 and 2009 as a control variable. Mian, Rao, and Sufi(2013) show that the deterioration in household balance sheets, asmeasured by county-level changes in household net worth, mighthave had a significant negative impact on local demand. Includ-ing this variable attenuates the sample size considerably, but ourmain results again remain unchanged. In summary, we have in-cluded a panoply of variables associated in the literature withthe household demand channel, and consistent with a credit sup-ply shock, Table III shows that nonbank-financed auto sales fellin those areas where borrowers were more heavily dependent onnonbank automotive credit.

V.B. Nonbank Dependence, Substitution, and Total Auto Sales

Here we examine the impact of nonbank dependence on salesfinanced by traditional deposit-taking institutions, as well as ontotal auto sales. If nonbank dependence proxies for latent demand,then a demand-side shock should induce a negative correlationbetween nonbank dependence and the growth in car sales, re-gardless of the lender’s source of funds. In contrast, a declinein the supply of nonbank credit could prompt banks and creditunions to increase their financing of automobiles in areas mostaffected by the loss of nonbank credit. When using the growth inbank-financed transactions as the dependent variable, this substi-tution would in turn lead to a positive coefficient on the nonbankdependence variable. Such a change in the sign of the coefficient



would be hard to reconcile with a “latent demand” interpretationof nonbank dependence.

Panel b of Online Appendix Table A.1 provides aggregate evi-dence consistent with substitution from nonbank credit to deposit-taking institutions during the crisis. The auto loan market shareof finance companies—the bulk of which is captive finance—was51.3% in 2005 and declined to just 41.3% and 36.7%, respectively,in 2009 and 2010. In contrast, the combined auto loan marketshare of credit unions and commercial banks rose from 44.9% in2005 to 56.2% and 61.1%, respectively, in 2009 and 2010.

We test the substitution hypothesis directly in Table IV. Weuse the same empirical specification as in column (1), but we re-define the dependent variable as the change in the number ofcars financed by banks and credit unions within a county be-tween 2008 and 2009. Similar to the analysis presented in Ta-ble III, standard errors (in parentheses) are clustered at the statelevel, and the regressions are weighted using county population.We use the same set of explanatory variables from column (2) ofTable III in the columns of Table IV.19 As Table IV, column (1)illustrates, the nonbank dependence point estimate is now posi-tive and statistically significant. A 1 standard deviation increasein captive dependence is associated with a 2.6 percentage point or0.15 standard deviation increase in bank-financed transactions inthe county. In Online Appendix Table A.7 we repeat the analysisin Table IV, column (1) for each of the specifications in Table IIIand find positive and statistically significant point estimates ofnonbank dependence in every specification. The evidence of par-tial substitution from nonbank providers of credit to traditionaldeposit-taking institutions lends additional credence to the creditsupply shock.

Next we analyze the aggregate consequences of the contrac-tion in nonbank credit supply. To do so, we redefine the dependentvariable as the log change in the number of all car sales in acounty between 2008 and 2009, regardless of whether they werefinanced or what the source of financing was. As Table IV, col-umn (2) demonstrates, the dependence coefficient is negative andstatistically significant. A 1 standard deviation increase in non-bank dependence is associated with a 1 percentage point or 0.08standard deviation decline in the growth in new car transactionsover this period.

19. For brevity, we do not report the coefficients on the socioeconomic anddemographic controls in Table IV.




TA

BL

EIV

NO

NB

AN

KD

EP

EN

DE

NC

EA

ND

AG

GR

EG

AT

EE

FF

EC

TS

Ban

k-fi

nan

ced

tran

sact

ion

sA

lltr

ansa

ctio

ns

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Var

iabl

es

Non

ban

kde

pen

den

ce:

Pol

k,20

06

Non

ban

kde

pen

den

ce:

Pol

k,20

06

Non

ban

kde

pen

den

ce:

Equ

ifax

,20

06

Non

ban

kde

pen

den

ce:

Equ

ifax

,20

08

Pol

k:E

xten

sive

mar

gin

Equ

ifax

cred

itsc

ore

Ch

ange

:20

09–2

007

Non

ban

kde

pen

den

ce,

0.25

7∗∗∗

−0.0

926∗

−0.0

940∗

−0.1

21∗

2008

Q1

(Pol

k),

fin

ance

dtr

ansa

ctio

ns

(0.0

831)

(0.0

490)

(0.0

503)

(0.0

660)

Non

ban

kde

pen

den

ce,

−0.1

17∗∗

∗E

quif

ax,2

006,

fin

ance

dtr

ansa

ctio

ns

(0.0

251)

Non

ban

kde

pen

den

ce,

−0.0

972∗

∗∗E

quif

ax,2

008,

Q1,

fin

ance

dtr

ansa

ctio

ns

(0.0

220)

Non

ban

kde

pen

den

ce,

−0.1

26∗∗

2008

Q1

(Pol

k),

exte

nsi

vem

argi

n(0

.056

7)


TA

BL

EIV

( CO

NT

INU

ED

)

Ban

k-fi

nan

ced

tran

sact

ion

sA

lltr

ansa

ctio

ns

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Var

iabl

es

Non

ban

kde

pen

den

ce:

Pol

k,20

06

Non

ban

kde

pen

den

ce:

Pol

k,20

06

Non

ban

kde

pen

den

ce:

Equ

ifax

,20

06

Non

ban

kde

pen

den

ce:

Equ

ifax

,20

08

Pol

k:E

xten

sive

mar

gin

Equ

ifax

cred

itsc

ore

Ch

ange

:20

09–2

007

Med

ian

cred

itsc

ore,

5.12

e−05

2008

Q1

(Equ

ifax

,n

onba

nks

)(5

.22e

−05)

Obs

erva

tion

s2,

825

2,82

52,

827

2,28

72,

825

2,49

42,

825

R2

0.70

00.

703

0.70

70.

723

0.70

50.

712

0.79

7

Not

es.T

his

tabl

ere

port

sre

gres

sion

resu

lts

ofes

tim

atin

geq

uat

ion

(1).

Th

ede

pen

den

tva

riab

lein

colu

mn

(1)

isth

elo

gch

ange

inth

en

um

ber

ofca

rsfi

nan

ced

byba

nks

and

cred

itu

nio

ns

in20

09re

lati

veto

2008

.Th

ede

pen

den

tva

riab

lein

colu

mn

s(2

)–(6

)is

the

log

chan

gein

the

tota

lnu

mbe

rca

rsso

ldin

aco

un

tyin

2009

rela

tive

to20

08,r

egar

dles

sof

the

sou

rce

offi

nan

cin

g.C

olu

mn

(7)

use

sth

elo

gch

ange

into

taln

um

ber

cars

sold

ina

cou

nty

in20

09re

lati

veto

2007

,reg

ardl

ess

ofth

eso

urc

eof

fin

anci

ng,

asth

ede

pen

den

tva

riab

le.C

olu

mn

(3)d

efin

esn

onba

nk

depe

nde

nce

asth

era

tio

ofn

onba

nk

tran

sact

ion

sto

allfi

nan

ced

tran

sact

ion

s,ba

sed

onda

tafr

omE

quif

axin

2006

.Col

um

n(4

)defi

nes

non

ban

kde

pen

den

ceas

the

rati

oof

non

ban

ktr

ansa

ctio

ns

toal

lfin

ance

dtr

ansa

ctio

ns,

base

don

data

from

Equ

ifax

in20

08Q

1.C

olu

mn

(5)d

efin

esn

onba

nk

depe

nde

nce

asth

era

tio

ofn

onba

nk

tran

sact

ion

sto

all

tran

sact

ion

s,ba

sed

onda

tafr

omE

quif

axin

2006

.Col

um

n(6

)ad

dsth

eE

quif

axm

easu

reof

med

ian

cred

itsc

ore

inth

eco

un

tyof

buye

rsu

sin

gn

onba

nk

sou

rces

ofau

tom

otiv

ecr

edit

,ob

serv

edin

2008

Q1,

toth

ese

tof

con

trol

vari

able

s.A

llco

lum

ns

incl

ude

the

sam

ese

tof

con

trol

sas

inT

able

III,

colu

mn

(2).

All

vari

able

sar

ede

fin

edin

App

endi

xA

.All

regr

essi

ons

are

wei

ghte

dby

the

cou

nty

popu

lati

onan

din

clu

dest

ate

fixe

def

fect

s.S

tan

dard

erro

rsar

ecl

ust

ered

atth

est

ate

leve

l.∗∗

∗ ,∗∗

,∗de

not

essi

gnifi

can

ceat

the

1%,5

%,a

nd

10%

leve

ls,

resp

ecti

vely

.



The implied economic impact of nonbank dependence on salesappears sizable. For each county we multiply its dependence onnonbank financing by the dependence coefficient from column (2).This product yields each county’s predicted growth in total carsales, as determined by the county’s degree of nonbank depen-dence. Multiplying this predicted growth rate by the level of salesin 2008 within the county gives the predicted change in car sales.Summing up across all counties suggests that the distress amongnonbanks might account for a drop of about 478,776 cars in 2009relative to 2008 sales; in our sample, 8.1 million cars were sold in2008 and 6.5 million in 2009. This implies that the liquidity shockto nonbank financing capacity can potentially explain 31% of thedrop in car sales in 2009 relative to 2008.

V.C. Measurement of Nonbank Dependence and Robustness Tests

We now check that our results are robust to alternative def-initions of the timing of the nonbank dependence measure. Ourbaseline Polk measure is calculated using 2008 Q1 data, whichmight already reflect some credit substitution from nonbank todepository institutions and hence might not fully represent de-pendence on nonbank entities within the county. We address thisconcern in Table IV, column (3). We use the same specification asin column (2) but use the Equifax-derived nonbank dependencemeasure, calculated using 2006 data, instead of the baseline Polkmeasure. This measure precedes the crisis, is computed over a fullyear, and likely measures nonbank dependence more precisely.

The point estimate in column (3) is negative and significantand is larger than the baseline estimate in column (2) (−0.117compared to −0.0926). A 1 standard deviation increase in thismeasure of nonbank dependence is associated with a 1.9 percent-age point or 0.16 standard deviation decline in the growth in totalcar sales. Computing nonbank dependence based on Equifax dataobserved in 2008 Q1 (column (4)) yields a point estimate that issimilar to that of column (2). In column (5) we use the extensivemargin measure of nonbank dependence from Polk—nonbank-financed transactions to all transactions—observed in 2008 Q1and find similar results. The results presented in Table IV demon-strate that the economic impact of the loss of nonbank financingcapacity appears significant and robust across various measuresof nonbank dependence from different sources.

Our baseline results are robust to the inclusion of the me-dian Transunion-based FICO score in the county, but to assuage


lingering measurement concerns, we turn again to Equifax micro-level data. Using these data, we calculate the median credit scorefor those borrowers who obtained nonbank automotive creditin the county in 2008 Q1, which allows us to measure moreaccurately the credit quality of nonbank customers. As column (6)shows, the point estimate on the baseline measure of dependenceis little changed, and the Equifax-derived measure of borrowercredit quality adds little information beyond the more generalTransunion credit quality variable.

We investigate the sensitivity of these results to the timingof the collapse in auto sales. Because the Polk data set does notidentify the source of credit before 2008 Q1, we have defined thecollapse in car sales as the change in sales between 2008 and2009. Given that the trough in annual sales occurred in 2009, thisapproach provides a reasonable approximation for the decline inauto sales (see Online Appendix Figure A.2). Nevertheless, dis-ruptions in the ABCP market started in late 2007, and car salesbegan falling in 2008. As a robustness exercise, we use the logdifference in car sales in 2009 relative to 2007 as the dependentvariable in Table IV, column (7). As the final column of the tabledemonstrates, the nonbank dependence coefficient is similar tothe one reported in column (2). A 1 standard deviation increase innonbank dependence is associated with a 1.3 percentage point or0.08 standard deviation decline in the growth in car sales, mea-sured between 2007 and 2009.

The panel structure of the data can also illuminate the timingof the collapse in car sales. To this end, we regress the quarterlygrowth in new car sales within a county from 2006 Q1 through2009 Q4. We include the baseline county-level controls along withour nonbank dependence measure. We interact the nonbank co-efficient with quarter dummies to permit the impact of nonbankdependence to vary by quarter over the sample period. These coef-ficients, along with the 95% confidence bands—in dashed lines—are plotted in Figure IV.

As the figure shows, in 2006, when nonbank lenders were ingeneral not financially constrained, car sales growth was positivein those counties where borrowers were more dependent on non-bank credit. The coefficient turns negative in the final quarter of2007 as the ABCP market became stressed and again in the quar-ters around the collapse of Lehman Brothers. Consistent with thenotion that the shutdown of the ABCP conduits of the major au-tomotive captive financing arms in early 2009 might have greatly



FIGURE IV

Captive Dependence and Car Sales, 2006 Q1–2009 Q4

The figure plots the coefficient (solid line) along with the 95% confidenceband (dashed line) from regressing the quarterly growth in aggregate car sales(at the county level) on captive dependence (Polk), and the baseline controls fromTable III, column (2), along with year-quarter fixed effects. The captive dependencecoefficient is allowed to vary by quarter over the sample period.

increased financing constraints among nonbank lenders, thecoefficient on nonbank dependence becomes even more negativein early 2009.

Changes in MMF flows—mutual funds that invest in short-term securities—over time can also shed light on the timing ofthe collapse in car sales (Online Appendix Figure A.4). BecauseMMFs were the principal source of funding for many securitiza-tion conduits, we would expect that when net flows into MMFsare plentiful, these funds are likely to increase their demand fornonbank automotive ABCP. Moreover, among MMFs, holdings ofABCP were highest among those that catered to institutional in-vestors and specialized in non-Treasury securities (Kacperczykand Schnabl 2013). In Online Appendix Table A.8, we show thatcar sales are more sensitive to aggregate fluctuations in short-term financing conditions—primarily flows into non-Treasury in-stitutional MMFs—in those counties more dependent on nonbankcredit.




In additional robustness tests, we repeat the main spec-ification in Table IV, column (2) for each of the four broadgeographic census regions and report results in Online AppendixTable A.9. Apart from the Northeast, where the small number ofobservations renders the estimates unreliable, the point estimateon nonbank dependence is negative across the regions, thoughimprecisely estimated when the sample size shrinks. Online Ap-pendix Tables A.10A, A.10B, and A.10C consider a battery of ad-ditional robustness tests. Online Appendix Table A.10A replicatesthe robustness exercises from Table III using total car purchasesas the dependent variable; the economic significance of the non-bank dependence coefficient is stable, but the coefficient becomesmore imprecise when the sample size declines and the regres-sions are weighted by population. Online Appendix Table A.10Bconducts the same exercise, but without the population weights;the nonbank dependence point estimate is negative and signifi-cant. Finally, Online Appendix Table A.10C repeats this exerciseusing the potentially more accurate nonbank dependence measurefrom Equifax (in 2006). Using the 2006 Equifax-based measure,we find that the point estimates are negative and robust acrossthe various specifications even with the smaller sample sizes andwhen weighted by population.

V.D. Make Heterogeneity and County Fixed Effects

We analyze the heterogeneity of the effect of nonbankdependence on auto sales across the three largest automanufacturers.20 In the first nine columns of Table V, we restrictour analysis to one automaker in each regression and estimatespecifications similar to regression (1) with the same set of con-trol variables as in Table III, column (2). Nonbank dependence isdefined as a county’s dependence on nonbank credit for each of theautomakers based on sales financed in 2008 Q1. Likewise, eachof the dependent variables is calculated using the correspondingauto sales of GM, Ford, or Toyota. The table reports results for thethree largest automakers in the United States: GM, columns (1)–(3); Ford, columns (4)–(6); and Toyota, columns (7)–(9).

The dependent variable in Table V, column (1) is the changein total GM sales within a county from 2008 to 2009. As the

20. There is evidence that concerns about the long-term solvency of the au-tomobile manufacturer could independently shape the demand for its cars (seeHortacsu et al. 2013).








TA

BL

EV

WIT

HIN

-MA

KE

EF

FE

CT

SO

FC

AP

TIV

EF

INA

NC

ING

ON

AU

TO

SA

LE

S

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Var

iabl

esA

llG

Msa

les

GM

AC

-fi

nan

ced

GM

Sal

es

Su

bsti

tuti

on:

non

-GM

AC

-fi

nan

ced

GM

sale

sA

llF

ord

sale

s

FM

C-

fin

ance

dF

ord

sale

s

Su

bsti

tuti

on:

non

-FM

C-

fin

ance

dF

ord

sale

s

All

Toy

ota

sale

s

TM

C-

fin

ance

dT

oyot

asa

les

Su

bsti

tuti

on:

non

-TM

C-

fin

ance

dT

oyot

asa

les

Cou

nty

,m

ake-

segm

ent

fixe

def

fect

s

Non

ban

k−0

.041

9∗−0

.425

∗∗∗

0.08

14∗∗

∗−0

.037

1∗∗−0

.146

∗∗∗

0.29

9∗∗∗

−0.0

202∗

−0.2

33∗∗

∗0.

192∗∗

∗−0

.026

2∗

depe

nde

nce

(0.0

228)

(0.0

537)

(0.0

288)

(0.0

182)

(0.0

266)

(0.0

295)

(0.0

117)

(0.0

613)

(0.0

264)

(0.0

145)

Obs

erva

tion

s2,

857

2,85

42,

857

2,85

72,

856

2,85

62,

855

2,83

72,

851

32,8

72R

20.

377

0.40

70.

509

0.38

90.

369

0.58

30.

289

0.32

80.

271

0.71

8

Not

es.T

he

depe

nde

nt

vari

able

inco

lum

n(1

)is

the

log

chan

gein

allG

Msa

les

in20

09re

lati

veto

2008

.Col

um

n(2

)is

the

log

chan

gein

GM

AC

-fin

ance

dG

Msa

les.

Col

um

n(3

)is

the

log

chan

gein

alln

on-G

MA

C-fi

nan

ced

GM

sale

s.C

olu

mn

s(4

)–(9

)fo

llow

asi

mil

arpa

tter

nfo

rF

ord

and

Toy

ota

sale

s.C

apti

vede

pen

den

ceis

defi

ned

asth

e20

08Q

1m

arke

tsh

ares

ofG

MA

C,F

MC

,an

dT

MC

,res

pect

ivel

y,in

apa

rtic

ula

rco

un

ty.I

nal

lca

ses,

the

shar

eof

the

mak

ein

tota

lco

un

tysa

les

isin

clu

ded

asa

regr

esso

ral

ong

wit

hth

ede

mog

raph

icco

ntr

ols

inT

able

III,

colu

mn

(2).

All

chan

ges

are

defi

ned

asth

epe

rcen

tage

chan

gein

2009

over

2008

.Col

um

n(1

0)st

acks

the

data

bym

ake:

GM

,For

d,T

oyot

a,an

dH

onda

;cou

nty

and

mod

else

gmen

t.C

olu

mn

(10)

incl

ude

sco

un

tyan

dbr

and

fixe

def

fect

s,al

ong

wit

hco

un

ty-s

egm

ent

fixe

def

fect

s.A

llre

gres

sion

sar

ew

eigh

ted

byth

eco

un

typo

pula

tion

and

incl

ude

stat

efi

xed

effe

cts.

Sta

nda

rder

rors

are

clu

ster

edat

the

stat

ele

vel.

∗∗∗ ,

∗∗,∗

den

otes

sign

ifica

nce

atth

e1%

,5%

,an

d10

%le

vels

,res

pect

ivel

y.


table shows, the point estimate on dependence is negative andsignificant, suggesting that the decline in GM sales was largerin those areas more dependent on GMAC: a 1 standard deviationincrease in dependence is associated with a 0.14 standard devi-ation drop in the change in GM-branded cars sales. Column (2)shows that although GM cars financed by GMAC fell in those ar-eas more dependent on GMAC, bank- and credit union–financedGM sales rose sharply in those areas in which nonbank providerslike GMAC were more dominant (column (3)).

The remaining columns of Table V repeat the basic specifi-cations for Ford and Toyota. The pattern is similar across thethree largest automakers. It suggests that despite the variationin experiences across these firms, dependence on captive financ-ing played a significant role in explaining a portion of the collapsein car sales.

Last, the richness of our data and particularly the availabilityof make- and model-level data allow us to once more gauge the ex-tent of biased estimates due to latent county-level unobservablesthat might explain the demand for cars within a county and itsdependence on captive financing. We build on the fact that the au-tomobile market is highly segmented and thus that shocks to thedemand for cars within a county could vary substantially acrossmodels, even for those sold by the same firm.

For example, some manufacturers, such as GM, offer a largenumber of makes and models aimed at buyers with different in-come levels: Chevrolet, a major make within GM, generally sellsnonluxury models that are marketed toward lower- and middle-income buyers, whereas Buick and Cadillac, also GM makes, sellmore luxurious models aimed at higher-income buyers.21 As a re-sult, the collapse in house prices and the rise in household lever-age among lower-income borrowers could precipitate a drop in thedemand for Chevrolet models within a county, whereas demandfor Buicks and Cadillacs within the same county could be lessaffected because house price dynamics might have had a smallerimpact on the net worth of these higher-income buyers.

Using the detailed model and make data from Polk, along withinformation on model types from Ward’s Automotive (one of the

21. Even within some makes such as Chevrolet, some models, like theCorvette, are aimed at richer buyers. Bricker, Ramcharan, and Krimmel (2014)and the references therein discuss cars, status, and the marketing of cars in theUnited States.


standard purveyors of intelligence on the automotive industry),we augment our analysis to employ within-make within-countywithin-segment heterogeneity. Ward’s identifies the market seg-ment in which each car model competes, and we use this infor-mation to construct a county-make-segment panel: the numberof cars that each make sold within each county in each marketsegment. The market segmentation in the industry can be highlydetailed, and Ward’s lists 30 segments. This level of granularitycan lead to a large number of missing observations in our dataset, as specialized models, such as the Chevrolet Corvette, tendto have a small number of sales in a limited geographic area. Wethus collapse the 30 segments in Ward’s into eight broad marketsegments that correspond to the Insurance Institute for HighwaySafety’s classification: small cars, mid-sized cars, large cars, lux-ury cars, small utility vehicles, mid-sized utility vehicles, largeutility vehicles, and luxury utility vehicles.22

With information on county, make, and segment, we can in-clude make fixed effects, county fixed effects, and county-segmentfixed effects. Make fixed effects allow us to absorb any shocks tomake-level sales that affects all counties and segments, such asthe potential insolvency of a make, while county fixed effects ab-sorb county-specific time-invariant factors that affect sales of allcars equally within the county. For example, a county’s exposureto the Cash for Clunkers program, as determined by the preexist-ing fraction of “clunkers” in the county’s automobile stock, couldbe correlated with sales in 2009 and with nonbank dependence(Mian and Sufi 2012). Similarly, a county’s industrial structure,such as the degree of employment in nontraded goods, or its indi-rect connections to the automobile sector not measured by Bureauof Labor Statistics employment shares, could also drive demandand correlate with the nonbank credit, leading to biased esti-mates. County-segment fixed effects, however, absorb invariantfactors that affect sales of a particular segment that vary acrosssegments, even within the same county.

As Table V, column (10) demonstrates, our basic results re-main the same when controlling for make and county-segmentfixed effects. A 1 standard deviation increase in captive depen-dence measured is associated with about a 1.2 percentage pointdrop in sales in 2009. In summary, the combined evidence in

22. Appendix B provides more details on how the Ward’s data are merged withPolk.



Table V renders it unlikely that our results are driven by omittedcounty or automaker factors. More important, column (10) showsthat our results hold when we compare cars sold within countyand auto segment, and thus it is unlikely that our captive depen-dence measure captures latent demand for cars.

VI. MICRO-LEVEL EVIDENCE FROM EQUIFAX

VI.A. Controlling for Individual Equifax Risk Scores

Our county-level analysis provides evidence that illiquidityamong nonbank lenders had a significant adverse effect on carsales. Although these results are robust to the inclusion of vari-ous measures of household demand and the housing cycle, thereremains a concern that county-level variation in nonbank depen-dence might reflect compositional differences in borrower creditquality and latent demand between nonbank and bank borrow-ers. Because of these differences, borrowers from captive leasingcompanies and other sources of nonbank credit might be morelikely to face a contraction in their credit limits imposed by otherlenders, such as credit card companies. Rather than reflecting theeffects of diminished nonbank financing induced by illiquidity inshort-term funding markets, these results might be an artifact ofa more general contraction in credit to risky borrowers.

To address these concerns, we turn to individual-level datafrom Equifax. Equifax records information about a person’sliabilities—automotive debt, mortgages, student loans, credit carddebt, and credit card borrowing limits—along with the person’sage, Equifax dynamic risk score, and zip code of residence. Inthe case of automotive debt, the data set also identifies whethercredit was obtained from a nonbank lender or a depository institu-tion. We use a 10% random sample from Equifax that we observequarterly from 2006 Q1 through 2009 Q4—a panel of about threemillion individuals.

These micro-level individual data enable us to study how ex-posure to nonbank financing—the degree of nonbank dependencein the county—might have affected an individual’s likelihood ofobtaining nonbank automotive credit, controlling directly for theborrower’s risk score, as well as other measures of borrower creditquality. We present summary statistics for the Equifax sample oncredit card balances, credit card limits, risk score, year of birth,and homeownership rate in Table VI for counties below and above


TABLE VINONBANK DEPENDENCE AND INDIVIDUAL-LEVEL CHARACTERISTICS, EQUIFAX

Credit cardbalance

Credit cardlimits

Equifaxrisk score

Year ofbirth

Homeownershiprate

Panel A: Counties below the median in “captive dependence”Mean 5,232 3,992.3 687.3 1957.4 0.18Median 4,301 3,774.3 690.4 1957.7 0.18Std. dev. 5,282 1,707.2 27.2 3.0 0.07

Panel B: Counties above the median in “captive dependence”Mean 5,594 3,439.5 674.2 1958.4 0.17Median 4,684 3,201.1 676.0 1958.6 0.17Std. dev. 4,772 1,980.1 29.4 3.0 0.07

Notes. This table reports summary statistics from the Equifax data set in 2007 at the county level. Panel arestricts the sample to individuals living in counties with a below median “captive dependence” share, definedas the ratio of captive financed transactions to all retail transactions, as recorded in Polk 2008 Q1. Panel brestricts the sample to those counties at or above the median level of captive dependence. Credit card balancesare computed for the sample of individuals with a positive balance.

the nonbank-dependence median. The Equifax-based summarystatistics in Table VI are consistent with the notion that countiesmore dependent on automotive nonbank finance generally havepopulations that register higher credit card balances and lowercredit limits, with concomitantly lower credit scores. The popula-tions in the more nonbank-dependent counties are also marginallyyounger and are less likely to own a home or at least havemortgage-related debt.

These potentially important differences in borrower composi-tion among users of nonbank credit render the household-leveltests even more important. By including the individual’s riskscore, age, and homeownership status, plus credit card balancesand revolving credit limits, we can directly control for key mea-sures of borrower credit quality. That is, unlike the more aggregatecounty-level evidence, these individual-level controls restrict thepotential for biased estimates that might arise from latent de-mand and unobserved differences in the composition of borrowersbetween nonbank lenders and other sources of automotive credit.The panel structure of these tests, which allow us to hold constantthese borrower-level observables and study how the variation innonbank financing capacity over the crisis period might have af-fected individual-level credit access, can offer powerful evidenceof the credit supply channel.

In Table VII, column (1) we use a linear probability modelto estimate the probability that an individual obtains nonbank

354 QUARTERLY JOURNAL OF ECONOMICST

AB

LE

VII

NO

NB

AN

KD

EP

EN

DE

NC

EA

ND

CA

RB

UY

ING,I

ND

IVID

UA

L-L

EV

EL

EV

IDE

NC

E

(1)

(2)

(3)

(4)

(5)

Non

ban

ksN

onba

nks

All

sale

sM

ortg

age

cred

itC

redi

tli

mit

Var

iabl

es20

08–2

009

2007

–200

920

08–2

009

2008

–200

920

08–2

009

Equ

ifax

risk

scor

e,lo

g−0

.015

9∗∗∗

−0.0

188∗∗

∗−0

.010

7∗∗∗

0.38

3∗∗∗

5.16

1∗∗∗

(0.0

0087

3)(0

.000

954)

(0.0

0074

3)(0

.014

7)(0

.240

)C

redi

tca

rdba

lan

ce,l

og−0

.000

329∗∗

∗−0

.000

358∗∗

∗−0

.000

838∗∗

∗0.

0052

4∗∗∗

0.92

3∗∗∗

(4.2

1e−0

5)(3

.66e

−05)

(7.7

2e−0

5)(0

.001

24)

(0.0

0328

)C

redi

tca

rdli

mit

,log

0.00

0308

∗∗∗

0.00

0335

∗∗∗

0.00

0872

∗∗∗

0.02

77∗∗

∗

(1.6

8e−0

5)(1

.69e

−05)

(3.8

1e−0

5)(0

.000

713)

Age

0.14

9∗∗∗

0.16

5∗∗∗

0.43

8∗∗∗

3.85

4∗∗∗

−46.

41∗∗

∗

(0.0

121)

(0.0

127)

(0.0

174)

(0.3

82)

(1.6

60)

Hom

eow

ner

indi

cato

r0.

0070

2∗∗∗

0.00

778∗∗

∗0.

0162

∗∗∗

2.11

9∗∗∗

(0.0

0028

8)(0

.000

318)

(0.0

0064

9)(0

.054

5)N

onba

nk

depe

nde

nce

(Pol

k)0.

0081

8∗∗∗

0.00

787∗∗

∗−0

.008

03∗

−0.0

112

−1.5

35∗∗

∗

(0.0

0190

)(0

.002

44)

(0.0

0415

)(0

.028

4)(0

.191

)N

onba

nk

depe

nde

nce

(Pol

k)∗

2007

Q2

3.41

e−05

(0.0

0075

9)N

onba

nk

depe

nde

nce

(Pol

k)∗

2007

Q3

0.00

133

(0.0

0260

)N

onba

nk

depe

nde

nce

(Pol

k)∗

2007

Q4

−0.0

0138

(0.0

0221

)N

onba

nk

depe

nde

nce

(Pol

k)∗

2008

Q1

−0.0

0062

9(0

.001

04)


TA

BL

EV

II(C

ON

TIN

UE

D)

(1)

(2)

(3)

(4)

(5)

Non

ban

ksN

onba

nks

All

sale

sM

ortg

age

cred

itC

redi

tli

mit

Var

iabl

es20

08–2

009

2007

–200

920

08–2

009

2008

–200

920

08–2

009

Non

ban

kde

pen

den

ce(P

olk)

∗20

08Q

2−0

.001

61∗∗

−0.0

0224

∗∗−0

.005

56∗∗

∗−0

.001

08−0

.075

7∗∗∗

(0.0

0065

3)(0

.001

03)

(0.0

0170

)(0

.002

30)

(0.0

175)

Non

ban

kde

pen

den

ce(P

olk)

∗20

08Q

3−0

.000

939

−0.0

0158

−0.0

0371

∗∗∗

−0.0

0281

−0.0

0740

(0.0

0068

5)(0

.001

17)

(0.0

0125

)(0

.002

45)

(0.0

417)

Non

ban

kde

pen

den

ce(P

olk)

∗20

08Q

4−0

.002

17∗∗

−0.0

0282

∗−0

.002

12−0

.000

755

0.06

19(0

.000

983)

(0.0

0157

)(0

.001

88)

(0.0

0270

)(0

.063

0)N

onba

nk

depe

nde

nce

(Pol

k)∗

2009

Q1

−0.0

0079

5−0

.001

480.

0016

20.

0017

30.

223∗∗

∗

(0.0

0127

)(0

.001

74)

(0.0

0192

)(0

.005

31)

(0.0

789)

Non

ban

kde

pen

den

ce(P

olk)

∗20

09Q

2−0

.001

47−0

.002

16−0

.002

510.

0007

280.

243∗∗

∗

(0.0

0127

)(0

.001

65)

(0.0

0224

)(0

.005

75)

(0.0

843)

Non

ban

kde

pen

den

ce(P

olk)

∗20

09Q

3−0

.006

64∗∗

∗−0

.007

33∗∗

∗−0

.009

98∗∗

∗−0

.000

442

0.22

9∗∗

(0.0

0180

)(0

.002

34)

(0.0

0320

)(0

.005

00)

(0.0

902)

Non

ban

kde

pen

den

ce(P

olk)

∗20

09Q

4−0

.002

50−0

.003

18−0

.002

51−0

.001

750.

226∗∗

∗

(0.0

0161

)(0

.002

16)

(0.0

0266

)(0

.006

24)

(0.0

820)

Obs

erva

tion

s23

,665

,802

35,5

61,9

5723

,665

,802

23,6

65,8

0223

,665

,802

R2

0.00

30.

003

0.00

60.

110

0.20

0

Not

es.T

he

depe

nde

nt

vari

able

inco

lum

ns

(1)

and

(2)

equ

als

1if

anin

divi

dual

fin

ance

da

car

purc

has

ein

the

quar

ter

thro

ugh

an

onba

nk

len

der

and

0ot

her

wis

e.T

he

depe

nde

nt

vari

able

inco

lum

n(3

)eq

ual

s1

ifan

indi

vidu

alfi

nan

ced

aca

rin

the

quar

ter,

rega

rdle

ssof

the

cred

itso

urc

ean

d0

oth

erw

ise.

Th

ede

pen

den

tva

riab

lein

colu

mn

(4)

equ

als

1if

the

indi

vidu

alob

tain

eda

mor

tgag

ein

the

quar

ter

and

0ot

her

wis

e.T

he

depe

nde

nt

vari

able

inco

lum

n(4

)is

the

log

ofth

ein

divi

dual

’scr

edit

lim

it.I

nal

lcol

um

ns,

the

data

are

quar

terl

yan

d,fo

rco

lum

ns

(1)

and

(3)–

(5),

are

obse

rved

from

2008

Q1–

2009

Q4;

colu

mn

(2)

incl

ude

sda

tafr

om20

07Q

1–20

09Q

4.


automotive credit in a given quarter in 2008–2009. Building onthe earlier panel-level results (Figure IV), which show that cap-tive financing capacity changed substantially over this period, weallow the coefficient on nonbank dependence at the county levelto vary by quarter. In addition to the household-level controls,we include state and year-by-quarter fixed effects, and we clusterstandard errors at the state level.

The evidence in column (1) suggests that with individual riskscore, age, credit card balance, and mortgage status held constant,individuals are more likely to obtain nonbank automotive creditwhen they live in a county with a greater dependence on nonbankcredit. Strikingly, however, the impact of nonbank dependence onthe probability of obtaining captive credit changes considerablyover the sample period. The coefficient drops by about 28% fromthe first quarter of 2008 to the final quarter of that year. It re-bounds slightly in the beginning of 2009 but drops sharply towardthe end of the year, by a factor of almost eight relative to its peakin the first quarter of 2008, and becomes insignificant in the thirdquarter of 2009.

To understand better the timing of the shock at the individuallevel, we extend the sample back to 2007. These results are re-ported in Table VII, column (2). Although turbulence in the hous-ing market and deleveraging already began in 2007, we find noevidence of a decline in the nonbank coefficient during this period(Mayer, Pence, and Sherlund 2009). Similar to the county-levelresults (Figure IV), the nonbank dependence coefficient becomessignificantly negative for the first time in the second quarter of2008. Also consistent with Figure IV, this point estimate remainsnegative throughout the remaining quarters. This suggests thatour findings are unlikely to be driven by omitted variables pertain-ing to the local housing market and its effects on consumer credit.In Online Appendix Table A.11A, column (1) we confirm thisresult by restricting the sample to 2007. These results are alsolittle changed if we model the persistence in car-buying behaviorwith a lagged dependent variable or if we control for borrower ob-servables using lagged values—observed either one quarter beforeor at the beginning of year.

Table VII, column (3) focuses on aggregate car sales over2008–2009. The dependent variable is the probability that an in-dividual obtains automotive credit, regardless of the source offinancing—excluding cash purchases, since Equifax has no infor-mation on these. As with column (1), for individuals living in a



more nonbank-dependent county, the likelihood of obtaining au-tomotive credit fell sharply at the end of 2009. In particular, thenonbank-dependence coefficient declines by about 33% in the thirdquarter of 2009 compared to its peak in the first quarter of 2008.This decline is less than the eightfold drop observed in column (1),as other sources of automotive financing might have substitutedfor the loss of nonbank financing.

We now consider a number of robustness tests. Table VI sug-gests that counties more dependent on nonbank finance might dif-fer from those more dependent on bank credit. To check whethernonbank dependence might more generally proxy for credit con-ditions within a county, Table VII, column (4) uses the probabilitythat the individual buys a home in the quarter as the depen-dent variable. If the nonbank-dependence variable reflects localcredit conditions, such as the supply of mortgage financing, thenthe nonbank-dependence coefficients should also evince a similarpattern to that observed in columns (1) and (2). The estimatesin column (4) show no such pattern. Instead, the nonbank depen-dence coefficient is insignificant. That said, the demand for housesmight have begun declining before the first quarter of 2008, andin Online Appendix Table A.11A, column (2) we repeat the anal-ysis for 2007–2009. The nonbank dependence coefficient remainsinsignificant.

To check further whether nonbank dependence might proxyfor other types of binding credit constraints at the individuallevel, Table VII, column (5) uses the log of the individual’s creditcard limit as the dependent variable.23 If anything, the nonbank-dependence point estimate becomes less negative and even posi-tive over time as the economy exited the recession in the secondhalf of 2009. As an alternative measure, we let the credit limitequal 1 for those households with positive limits and 0 otherwise.We also consider the credit utilization rate— the ratio of credit bal-ances to limits—among those individuals with positive credit lim-its. In both cases, there is no evidence that nonbank dependencemight proxy for other types of binding credit constraints as demon-strated in Online Appendix Table A.11A, columns (3) and (4).

Last, we aggregate the individual-level data up to thecounty to check further whether nonbank dependence might beassociated with broader credit outcomes. That is, we examinewhether the change in mortgage balances between 2008 and 2009

23. We transform the credit card limit to log(1 + credit card limit).




within a county, or the log level of mortgage balances in 2009,is correlated with the county’s nonbank dependence. Similarly,we aggregate up credit card balances and examine whether thechange in the total credit card balance between 2008 and 2009at the county level is correlated with nonbank dependence. In allcases the nonbank dependence coefficient is not significant. Theseresults are presented in Online Appendix Table A.11B.

VI.B. Captive Dependence and Local Auto Sales Stratifiedby Equifax Risk Score

Reputational motives as well as declining collateral valuescan prompt financial institutions to tighten credit policies afteran adverse shock.24 Therefore, to gauge further the robustnessof our results and understand better the underlying channelsthrough which the financing shock might have led to the drop incar sales, we examine how the impact of exposure to nonbank fi-nancing on the likelihood of obtaining nonbank automotive creditmight have varied by borrower credit quality. To this end, weestimate the baseline specification in Table VII, column (1) sep-arately for borrowers with different risk scores and report theresults in Table VIII. We stratify the Equifax data by risk scorequartiles: column (1) uses the subsample of borrowers in the low-est quartile—those with a risk score below 603; column (2) usesborrowers from the second quartile, between 603 and 706; column(3) focuses on the third quartile, 706–784; and column (4) containsthose borrowers with scores above 784.

Across all borrower risk score categories, the point estimatesimply that access to nonbank automotive credit declined sharplytoward the end of 2008 and again in the second half of 2009. Forexample, even among those borrowers in the top quartile, the non-bank dependence coefficient, although positive in the first quarterof 2008, declines by about 43% in the third quarter of 2009 rela-tive to its value in the first quarter of 2008. But consistent withthe idea that credit policy might have become especially conser-vative after a shock, the decline in nonbank credit access appearsto be most severe for those borrowers with risk scores in the bot-tom quartile. From column (1) the overall impact of dependencein the third quarter of 2009 is negative, suggesting that these

24. See Bernanke and Gertler (1989), Rajan (1994), and the loan-level evidencefrom the financial crisis in Ramcharan, van den Heuvel, and Verani (2016) andRamcharan and Crowe (2013).



TABLE VIIICAPTIVE DEPENDENCE AND CAR BUYING STRATIFED BY EQUIFAX RISK SCORE

Variables Quartile 1 Quartile 2 Quartile 3 Quartile 4

Captive 0.00290 0.0129∗∗∗ 0.00980∗∗∗ 0.00739∗∗∗dependence(Polk)

(0.00378) (0.00223) (0.00144) (0.00210)

Captive −0.00214 −0.00184 5.59e−06 −0.00244dependence(Polk) ∗ 2008 Q2

(0.00272) (0.00166) (0.00128) (0.00194)

Captive −0.00252 −0.000426 −0.000534 0.000776dependence(Polk) ∗ 2008 Q3

(0.00232) (0.00146) (0.00113) (0.000887)

Captive −0.00175 −0.000221 −0.00261∗∗ −0.00254∗∗dependence (Polk)

∗ 2008 Q4(0.00251) (0.00159) (0.00127) (0.00118)

Captive 0.000400 −0.000336 −0.000574 −0.000840dependence(Polk) ∗ 2009 Q1

(0.00270) (0.00173) (0.00138) (0.00113)

Captive 0.000407 −0.000803 −0.000461 −0.00287∗dependence(Polk) ∗ 2009 Q2

(0.00332) (0.00184) (0.00105) (0.00154)

Captive −0.00557∗∗ −0.00847∗∗∗ −0.00576∗∗∗ −0.00342∗dependence(Polk) ∗ 2009 Q3

(0.00269) (0.00273) (0.00199) (0.00178)

Captive −0.00208 −0.00308 −0.000748 −0.00140dependence(Polk) ∗ 2009 Q4

(0.00328) (0.00188) (0.00174) (0.00146)

Observations 5,870,586 5,934,242 5,896,749 5,858,620R2 0.002 0.003 0.003 0.003

Notes. The dependent variable equals 1 if an individual financed a car purchase in the quarter through anonbank and 0 otherwise. All columns include the same controls as in Table VII; observed from 2008 Q1–2009Q4, and standard errors are clustered at the state level. Column (1) includes individuals with a risk scorebelow 603; column (2) uses scores between 603 and 706 (second quartile); column (3) uses 706–784 (thirdquartile); and column (4) focuses on individuals with scores above 784.

borrowers were less likely to obtain captive credit in those areasmore dependent on nonbank financing.

VII. CONCLUSION

There is growing consensus that the financial crisis of 2008–2009 shared many elements of a bank run, but one concentratedin the non-deposit-taking sectors of the financial system thatprimarily relied on short-term funding. Less well understood,however, are the economic consequences of the runs in these fund-ing markets. Focusing on the U.S. automobile sector, we show that


the illiquidity in short-term funding markets during the crisismight have played an important role in limiting the supply ofnonbank automotive credit, such as automotive captive lenders,during the financial crisis. In particular, our estimates suggestthat this contraction in the supply of nonbank automotive credit,largely due to the illiquidity in the ABCP market, might explainabout one-third of the collapse in car sales during this period,possibly further worsening the financial situation of the majorU.S. automakers.

This evidence is related to the broader literature on how fund-ing disruptions and other balance sheet shocks to traditional fi-nancial institutions might affect credit availability to the realeconomy. Our results also suggest that although shocks to thebalance sheet of households might account for a substantial partof the decline in economic activity after the crisis, illiquidity inshort-term funding markets and balance shocks to both bank andnonbank institutions might also explain some of this decline, de-spite myriad policy interventions. We leave it to future researchto measure more precisely the efficacy of these interventions.

APPENDIX

A: Variable Description and Construction

For reference, the following is a list of variables used in thearticle, their sources, and a brief description of how each variableis constructed.

African American population: Number of African Americans in acounty. (Source: American Community Survey)

Assets: Total bank assets. (Source: FR Y9-C, FFIEC 031)Captive dependence: Share of county-level retail car sales financed

by captive financing companies. (Source: Polk)Captive financed sales: County-level retail car sales financed by

captive financing companies. (Source: Polk)County area: Size of a county in square miles. (Source: American

Community Survey)Employment in automobile manufacturing: Divides the number

of employees in the automobile sector by total employment.(Source: Quarterly Census of Employment and Wages)

Gini coefficient: Measures income inequality in a county. (Source:American Community Survey)


House price change: Annual change in the local house price index.(Source: CoreLogic)

Household leverage: County-level household debt-to-income ratio.(Source: Federal Reserve of New York)

Median household income (Source: American Community Survey)Median credit score, 2008 Q1 (Trans Union): The median FICO

score in the county in 2008 Q1 from Trans Union, drawnfrom the entire population in the county.

Median credit score, 2008 Q1 (Equifax): The median Equifax riskscore in the county in 2008 Q1 among buyers using captivesfor automotive credit.

Money market fund flows: Quarterly net flows to (from) moneymarket funds. (Source: Flow of Funds, Federal Reserve Board)

Noncaptive financed sales: County-level retail car sales not fi-nanced by captive financing companies. (Source: Polk)

Percent African American: African American population dividedby total population. (Source: American Community Survey)

Population: Number of people in a county. (Source: American Com-munity Survey)

Population density: Population divided by area. (Source: AmericanCommunity Survey)

Poverty rate: Number of people living below the poverty line di-vided by population. (Source: U.S. Census)

Retail car sales: The sum of retail purchases and retail leases.(Source: Polk)

Unemployment rate: County-level labor force divided by the num-ber of unemployed. (Source: Bureau of Labor Statistics)

White population: Number of Caucasians in a county. (Source:American Community Survey)

B: Auto Segment Construction

The eight auto segments used in make-county regression(Table V) include the following models:

Small cars (Ward’s categories: lower small and upper small): BMW128, BMW 135, Chevrolet Aveo, Chevrolet Cobalt, DodgeCaliber, Ford Focus, Honda Civic, Honda Fit, Hyundai Ac-cent, Hyundai Elantra, Kia Forte, Kia Rio, Kia Soul, Kia Spec-tra, Mazda 3, Mini Cooper, Mitsubishi Lancer, Nissan Cube,


Nissan Sentra, Nissan Versa, Pontiac G3, Pontiac Vibe, Saab93, Saturn Astra, Subaru Impreza, Saturn Ion, Suzuki Aerio,Suzuki Forenza, Suzuki Reno, Suzuki SX4, Toyota Corolla,Toyota Yaris, Volkswagen GLI, Volkswagen Golf, VolkswagenJetta, Volkswagen Rabbit, Volkswagen R32, Volvo V50.

Mid-sized cars (Ward’s categories: lower middle and upper mid-dle): Buick Lacrosse, Chevrolet Impala, Chevrolet Malibu,Chrysler Sebring, Dodge Avenger, Ford Fusion, Honda Ac-cord, Honda FCX, Honda Insight, Hyundai Azera, HyundaiSonata, Kia Optima, Mazda 6, Mercury Mila, Mercury Mon-tego, Mercury Sable, Mitsubishi Galant, Nissan Altima, Pon-tiac G6, Pontiac G8, Pontiac Grand Prix, Saturn Aura, Sub-aru Legacy, Suzuki Kizashi, Toyota Camry, Volkswagen CC,Volkswagen Passat, Volvo V70.

Large cars (Ward’s category: large): Buick Lucerne, Chrysler 300,Dodge Charger, Dodge Magnum, Ford Crown Victoria, FordFive Hundred, Ford Taurus, Kia Amanti, Mercury GrandMarquis, Mercury Monterey.

Luxury cars (Ward’s categories: small luxury, middle luxury, andlarge luxury): Acura RL, Acura TL, Acura TSX, Audi A3,Audi A4, Audi A6, Audi S4, Bentley Continental, BMW 328,BWM 335, BW 525, BMW 528, BMW 530, BMW 535, BMW550, BMW M3, BMW M5, Cadillac CTS, Cadillac DTS, Cadil-lac STS, Chevrolet Monte Carlo, Hyundai Genesis, InfinitiG35, Infiniti G37, Infiniti M35, Infiniti M45, Jaguar S-Type,Jaguar X-Type, Lexus ES, Lexus GS, Lexus HS250H, LexusIS, Lincoln MKS, Lincoln MKZ, Lincoln Town Car, Mercedes-Benz C-Class, Mercedes-Benz CLK-Class, Mercedes-Benz E-Class, Nissan Maxima, Toyota Avalon, Volvo S40, Volvo S60,Volvo S80.

Small utility vehicles (Ward’s categories: small cross/utility andsmall sport/utility): Chevrolet HHR, Chrysler PT Cruiser,Dodge Nitro, Honda Element, Hyundai Tucson, Jeep Com-pass, Jeep Liberty, Jeep Patriot, Jeep Wrangler, Kia Sportage,Land Rover LR2, Mercury Mariner, Saab 95, Suzuki GrandVitara.

Mid-sized utility vehicles (Ward’s categories: middle cross/utilityand middle sport/utility): Chevrolet Equinox, ChevroletTrailblazer, Dodge Journey, Ford Edge, Ford Escape, FordExplorer, GMC Envoy, GMC Terrain, Honda CR-V, Honda


Crosstour, Honda Pilot, Hyundai Santa Fe, Hyundai Ver-acruz, Isuzu Ascender, Jeep Commander, Jeep Grand Chero-kee, Kia Borrego, Kia Rondo, Kia Sorento, Land Rover LR3,Mazda CX-7, Mazda 5, Mazda Tribute, Mitsubishi Endeavor,Mitsubishi Outlander, Nissan Murano, Nissan Pathfinder,Nissan Rogue, Nissan Xterra, Pontiac Torrent, Saturn Vue,Subaru B9 Tribeca, Subaru Forester, Subaru Outback, SuzukiXL7, Toyota 4 Runner, Toyota FJ Cruiser, Toyota Highlander,Toyota RAV4, Toyota Venza, Volkswagen Tiguan.

Large utility vehicles (Ward’s categories: large cross/utility andlarge sport/utility): Buick Enclave, Chevrolet Suburban,Chevrolet Tahoe, Chevrolet Traverse, Chrysler Aspen, DodgeDurango, Ford Expedition, Ford Flex, Ford Freestyle, FordTaurus X, GMC Acadia, GMC Envoy XL, GMC Yukon, MazdaCX-9, Mitsubishi Montero, Nissan Armada, Saturn Outlook,Toyota Sequoia.

Luxury utility vehicles (Ward’s categories: small luxurycross/utility, middle luxury cross/utility, large luxurycross/utility, luxury middle sport/utility, and luxury largesport/utility): Acura MDX, Acura RDX, Acura ZDX, AudiQ5, Audi Q7, BMW X3, BMW X5, BMW X6, Buick Rainier,Buick Rendezvous, Cadillac Escalade, Cadillac SRX, ChryslerPacifica, Hummer 4-PSGR Wagon, Hummer H2, HummerH3, Infiniti EX, Infiniti FX35, Infiniti FX45, Infiniti FX50,Infiniti QX56, Land Rover LR4, Land Rover Range Rover,Lexus GX, Lexus LX, Lexus RX, Lincoln MKT, LincolnMKX, Lincoln Navigator, Mercedes-Benz G-class, Mercedes-Benz GL-class, Mercedes-Benz GLK, Mercedes-Benz M-class, Mercedes-Benz R-class, Mercury Mountaineer, PorscheCayenne, Saab 9-7X, Subaru Tribeca, Toyota Land Cruiser,Volkswagen Touareg, Volvo XC60, Volvo XC70, Volvo XC90.

NORTHWESTERN UNIVERSITY AND NBERFEDERAL RESERVE

UNIVERSITY OF SOUTHERN CALIFORNIA

SUPPLEMENTARY MATERIAL

An Online Appendix for this article can be found at The Quar-terly Journal of Economics online.




REFERENCES

Acharya, Viral V., and Nada Mora, “A Crisis of Banks as Liquidity Providers,”Working Paper, 2013.

Acharya, Viral V., Phillip Schnabl, and Gustavo Suarez, “Securitization withoutRisk Transfer,” Journal of Financial Economics, 107 (2011), 515–536.

Adams, William, Liran Einav, and Jonathan Levin, “Liquidity Constraints andImperfect Information in Subprime Lending,” American Economic Review, 99(2009), 49–84.

Autor, David H., and David Dorn, “The Growth of Low-Skilled Service Jobs andthe Polarization of the US Labor Market,” American Economic Review, 103(2013), 1553–1597.

Barron, John M., Byung-Un Chong, and Michael E. Staten, “Emergence of CaptiveFinance Companies and Risk Segmentation in Loan Markets: Theory andEvidence,” Journal of Money, Credit and Banking, 40 (2008), 173–192.

Becker, Bo, and Victoria Ivashina, “Cyclicality of Credit Supply: Firm Level Evi-dence,” Journal of Monetary Economics, 62 (2014), 76–93.

Bernanke, Ben, and Mark Gertler, “Agency Costs, Net Worth and Business Fluc-tuations,” American Economics Review, 79 (1989), 14–31.

Bricker, Jesse, Rodney Ramcharan, and Jacob Krimmel, “Signaling Sta-tus: The Impact of Relative Income on Household Consumptionand Financial Decisions,” FEDS Working Paper no. 2014-76, 2014,http://dx.doi.org/10.2139/ssrn.2435503.

Brunnermeier, Markus K., “Deciphering the Liquidity and Credit Crunch 2007–2008,” Journal of Economic Perspectives, 23 (2009), 77–100.

Calder, Lendol, Financing the American Dream: A Cultural History of ConsumerCredit (Princeton, NJ: Princeton University Press, 1999).

Campbell, Sean, Daniel Covitz, William Nelson, and Karen Pence, “Securitiza-tion Markets and Central Banking: An Evaluation of the Term Asset-BackedSecurities Loan Facility,” Journal of Monetary Economics, 58 (2011), 518–531.

Charles, Kerwin Kofi, Erik Hurst, and Melvin Stephens, “Rates for Vehicle Loans:Race and Loan Source,” American Economic Review, 98 (2010), 315–320.

Chodorow-Reich, Gabriel, “The Employment Effects of Credit Market Disruptions:Firm-Level Evidence from the 2008–9 Financial Crisis,” Quarterly Journal ofEconomics, 129 (2014), 1–59.

Cornett, Marcia Millon, Jamie John McNutt, Philip E. Strahan, and Hassan Tehra-nian, “Liquidity Risk Management and Credit Supply in the Financial Crisis,”Journal of Financial Economics, 101 (2011), 297–312.

Covitz, Daniel, Nellie Liang, and Gustavo A. Suarez, “The Evolution of a Finan-cial Crisis: Collapse of Asset-Backed Commercial Paper Market,” Journal ofFinance, 68 (2013), 815–848.

Diamond, Douglas W., and Raghuram G. Rajan, “Liquidity Shortages and BankingCrisis,” Journal of Finance, 60 (2005), 615–647.

———, “Fear of Fire Sales, Illiquidity Seeking, and Credit Freezes,” QuarterlyJournal of Economics, 126 (2011), 557–591.

DiMaggio, Marco, Amir Kermani, and Rodney Ramcharan, “Monetary Policy Pass-Through: Household Consumption and Voluntary Deleveraging,” Working Pa-per, Columbia Business School, 2014.

Einav, Liran, Mark Jenkins, and Jonathan Levin, “Contract Pricing in ConsumerCredit Markets,” Econometrica, 80 (2012), 1387–1432.

———, “The Impact of Credit Scoring on Consumer Credit,” RAND Journal ofEconomics, 44 (2013), 249–274.

Gao, Pengjie, and Hayong Yun, “Liquidity Backstop, Corporate Borrowing, andReal Effects,” Working Paper, 2013.

Goolsbee, Austan, and Allan Krueger, “A Retrospective Look at Rescuing and Re-structuring General Motors and Chrysler,” Journal of Economic Perspectives,29 (2015), 3–24.

Gorton, Gary B., Slapped by the Invisible Hand: The Panic of 2007 (Oxford: OxfordUniversity Press, 2010).

http://dx.doi.org/10.2139/ssrn.2435503


Gorton, Gary B., and Andrew Metrick, “Securitized Banking and the Run on Repo,”Journal of Financial Economics, 104 (2012), 425–451.

Hortacsu, Ali, Gregor Matvos, Chad Syverson, and Sriram Venkataraman, “In-direct Costs of Financial Distress in Durable Goods Industries: The Case ofAuto Manufacturers,” Review of Financial Studies, 26 (2013), 1248–1290.

Hyman, Louis, Debtor Nation: The History of America in Red Ink (Princeton, NJ:Princeton University Press, 2011).

Ivanshina, Victoria, and David Scharfstein, “Bank Lending during the FinancialCrisis of 2008,” Journal of Financial Economics, 97 (2010), 319–338.

Kacperczyk, Marcin, and Philipp Schnabl, “When Safe Proved Risky: Commer-cial Paper during the Financial Crisis of 2007–2009,” Journal of EconomicPerspectives, 24 (2010), 29–50.

———, “How Safe Are Money Market Funds?” Quarterly Journal of Economics,128 (2013), 1073–1122.

Kaisha, Jidosha, Toyota: A History of the First 50 Years (Toyota City: Toyota MotorCorporation, 1988).

Khwaja, Asim Ijaz, and Atif Mian, “Tracing the Impact of Bank Liquidity Shocks:Evidence from an Emerging Market,” American Economic Review, 98 (2008),1413–1442.

Ludvigson, Sydney, “The Channel of Monetary Transmission to Demand: Evidencefrom the Market for Automobile Credit,” Journal of Monetary Economics, 30(1998), 365–383.

Mayer, Christopher J., Karen M. Pence, and Shane M. Sherlund, “The Rise inMortgage Defaults,” Journal of Economic Perspectives, 23 (2009), 27–50.

Mian, Atif, Kamalesh Rao, and Amir Sufi, “Household Balance Sheets, Consump-tion and the Economic Slump,” Quarterly Journal of Economics, 128 (2013),1687–1726.

Mian, Atif, and Amir Sufi, “Household Leverage and the Recession of 2007–09,”IMF Economic Review, 58 (2010), 74–114.

———, “House Prices, Home Equity-Based Borrowing, and the U.S. HouseholdLeverage Crisis,” American Economic Review, 101 (2011), 2132–2156.

———, “The Effects of the Fiscal Stimulus: Evidence from the 2009 Cash forClunkers Program,” Quarterly Journal of Economics, 127 (2012), 1107–1142.

———, “What Explains the 2007–2009 Drop in Employment? The Aggregate De-mand Channel,” Working Paper, Princeton University, 2014a.

———, House of Debt: How They (and You) Caused the Great Recession, and HowWe Can Prevent It from Happening Again (Chicago: University of ChicagoPress, 2014b).

Murfin, Justin, and Ryan Pratt, “Who Finances Durable Goods and Why It Mat-ters: Captive Finance and the Coase Conjecture,” Working Paper, Yale Schoolof Management, 2015.

Phelps, Clyde William, The Role of the Sales Finance Companies in the AmericanEconomy (Baltimore: Commercial Credit, 1952).

Rajan, Raghuram, “Why Bank Credit Policies Fluctuate: A Theory and Some Evi-dence,” Quarterly Journal of Economics, 109 (1994), 399–441.

Rajan, Raghuram, and Rodney Ramcharan, “The Anatomy of a Credit Crisis: TheBoom and Bust in Farm Land Prices in the United States in the 1920s,”American Economic Review, 105 (2015), 1439–1477.

———, “Local Financial Capacity and Asset Values: Evidence from Bank Failures,”Journal of Financial Economics, forthcoming.

Ramcharan, Rodney, and Christopher Crowe, “The Impact of House Prices onConsumer Credit: Evidence from an Internet Bank,” Journal of Money, Creditand Banking, 45 (2013), 1085–1115.

Ramcharan, Rodney, Skander J. van den Heuvel, and Stephane Verani, “FromWall Street to Main Street: The Impact of the Financial Crisis on ConsumerCredit,” Journal of Finance, 71 (2016), 1323–1356 .

Shleifer, Andrei, and Robert W. Vishny, “Unstable Banking,” Journal of FinancialEconomics, 97 (2010), 306–318.

Sloan, Alfred, My Years with General Motors (New York: Doubleday, 1964).

Documents

THE REAL EFFECTS OF LIQUIDITY DURING THE FINANCIAL … · General Motors Acceptance Corporation (GMAC), the ﬁnancing arm of GM and one of the world’s largest providers of auto